Matrix Analysis

Matrix Analysis
for Scientists & Engineers

Matrix Analysis
This page intentionally left blank This page intentionally left blank
Matrix Analysis
Alan J. Laub
University of California
Davis, California
slam.
Matrix Analysis
Alan J. Laub
University of California
Davis, California
Copyright 2005 by the Society for Industrial and Applied Mathematics.
1 0 9 8 7 6 5 4 3 2 1
All rights reserved. Printed in the United States of America. No part of this book
may be reproduced, stored, or transmitted in any manner without the written permission
of the publisher. For information, write to the Society for Industrial and Applied
Mathematics, 3600 University City Science Center, Philadelphia, PA 19104-2688.
MATLAB is a registered trademark of The MathWorks, Inc. For MATLAB product information,
please contact The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA 01760-2098 USA,
508-647-7000, Fax: 508-647-7101, info@mathworks.com, www.mathworks.com
Mathematica is a registered trademark of Wolfram Research, Inc.
Mathcad is a registered trademark of Mathsoft Engineering & Education, Inc.
Library of Congress Cataloging-in-Publication Data
Laub, Alan J., 1948-
Matrix analysis for scientists and engineers / Alan J. Laub.
p. cm.
Includes bibliographical references and index.
ISBN 0-89871-576-8 (pbk.)
1. Matrices. 2. Mathematical analysis. I. Title.
QA188138 2005
512.9'434dc22
2004059962
About the cover: The original artwork featured on the cover was created by freelance
artist Aaron Tallon of Philadelphia, PA. Used by permission.
slam is a registered trademark.
Copyright 2005 by the Society for Industrial and Applied Mathematics.
10987654321
All rights reserved. Printed in the United States of America. No part of this book
may be reproduced, stored, or transmitted in any manner without the written permission
of the publisher. For information, write to the Society for Industrial and Applied
Mathematics, 3600 University City Science Center, Philadelphia, PA 19104-2688.
MATLAB is a registered trademark of The MathWorks, Inc. For MATLAB product information,
please contact The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA 01760-2098 USA,
508-647-7000, Fax: 508-647-7101, info@mathworks.com, wwwmathworks.com
Mathematica is a registered trademark of Wolfram Research, Inc.
Mathcad is a registered trademark of Mathsoft Engineering & Education, Inc.
Library of Congress Cataloging-in-Publication Data
Laub, Alan J., 1948-
Matrix analysis for scientists and engineers / Alan J. Laub.
p. cm.
Includes bibliographical references and index.
ISBN 0-89871-576-8 (pbk.)
1. Matrices. 2. Mathematical analysis. I. Title.
QA 188.L38 2005
512.9'434-dc22
2004059962
About the cover: The original artwork featured on the cover was created by freelance
artist Aaron Tallon of Philadelphia, PA. Used by permission .

5.lam... is a registered trademark.
To my wife, Beverley
(who captivated me in the UBC math library
nearly forty years ago)
To my wife, Beverley
(who captivated me in the UBC math library
nearly forty years ago)
Contents
Preface xi
1 Introduction and Review 1
1.1 Some Notation and Terminology 1
1.2 Matrix Arithmetic 3
1.3 Inner Products and Orthogonality 4
1.4 Determinants 4
2 Vector Spaces 7
2.1 Definitions and Examples 7
2.2 Subspaces 9
2.3 Linear Independence 10
2.4 Sums and Intersections of Subspaces 13
3 Linear Transformations 17
3.1 Definition and Examples 17
3.2 Matrix Representation of Linear Transformations 18
3.3 Composition of Transformations 19
3.4 Structure of Linear Transformations 20
3.5 Four Fundamental Subspaces 22
4 Introduction to the Moore-Penrose Pseudoinverse 29
4.1 Definitions and Characterizations 29
4.2 Examples 30
4.3 Properties and Applications 31
5 Introduction to the Singular Value Decomposition 35
5.1 The Fundamental Theorem 35
5.2 Some Basic Properties 38
5.3 Row and Column Compressions 40
6 Linear Equations 43
6.1 Vector Linear Equations 43
6.2 Matrix Linear Equations 44
6.3 A More General Matrix Linear Equation 47
6.4 Some Useful and Interesting Inverses 47
vii
Contents
Preface
1 Introduction and Review
1.1 Some Notation and Terminology
1.2 Matrix Arithmetic . . . . . . . .
1.3 Inner Products and Orthogonality .
1.4 Determinants
2 Vector Spaces
2.1 Definitions and Examples .
2.2 Subspaces.........
2.3 Linear Independence . . .
2.4 Sums and Intersections of Subspaces
3 Linear Transformations
3.1 Definition and Examples . . . . . . . . . . . . .
3.2 Matrix Representation of Linear Transformations
3.3 Composition of Transformations . .
3.4 Structure of Linear Transformations
3.5 Four Fundamental Subspaces . . . .
4 Introduction to the Moore-Penrose Pseudoinverse
4.1 Definitions and Characterizations.
4.2 Examples..........
4.3 Properties and Applications . . . .
5 Introduction to the Singular Value Decomposition
5.1 The Fundamental Theorem . . .
5.2 Some Basic Properties .....
5.3 Rowand Column Compressions
6 Linear Equations
6.1 Vector Linear Equations . . . . . . . . .
6.2 Matrix Linear Equations ....... .
6.3 A More General Matrix Linear Equation
6.4 Some Useful and Interesting Inverses.
vii
xi
1
1
3
4
4
7
7
9
10
13
17
17
18
19
20
22
29
29
30
31
35
35
38
40
43
43
44
47
47
viii Contents
7 Projections, Inner Product Spaces, and Norms 51
7.1 Projections 51
7.1.1 The four fundamental orthogonal projections 52
7.2 Inner Product Spaces 54
7.3 Vector Norms 57
7.4 Matrix Norms 59
8 Linear Least Squares Problems 65
8.1 The Linear Least Squares Problem 65
8.2 Geometric Solution 67
8.3 Linear Regression and Other Linear Least Squares Problems 67
8.3.1 Example: Linear regression 67
8.3.2 Other least squares problems 69
8.4 Least Squares and Singular Value Decomposition 70
8.5 Least Squares and QR Factorization 71
9 Eigenvalues and Eigenvectors 75
9.1 Fundamental Definitions and Properties 75
9.2 Jordan Canonical Form 82
9.3 Determination of the JCF 85
9.3.1 Theoretical computation 86
9.3.2 On the +1's in JCF blocks 88
9.4 Geometric Aspects of the JCF 89
9.5 The Matrix Sign Function 91
10 Canonical Forms 95
10.1 Some Basic Canonical Forms 95
10.2 Definite Matrices 99
10.3 Equivalence Transformations and Congruence 102
10.3.1 Block matrices and definiteness 104
10.4 Rational Canonical Form 104
11 Linear Differential and Difference Equations 109
11.1 Differential Equations 109
11.1.1 Properties of the matrix exponential 109
11.1.2 Homogeneous linear differential equations 112
11.1.3 Inhomogeneous linear differential equations 112
11.1.4 Linear matrix differential equations 113
11.1.5 Modal decompositions 114
11.1.6 Computation of the matrix exponential 114
11.2 Difference Equations 118
11.2.1 Homogeneous linear difference equations 118
11.2.2 Inhomogeneous linear difference equations 118
11.2.3 Computation of matrix powers 119
11.3 Higher-Order Equations 120
viii
7 Projections, Inner Product Spaces, and Norms
7.1 Projections ..................... .
7.1.1 The four fundamental orthogonal projections
7.2 Inner Product Spaces
7.3 Vector Norms
7.4 Matrix Norms ....
8 Linear Least Squares Problems
8.1 The Linear Least Squares Problem . . . . . . . . . . . . . .
8.2 Geometric Solution . . . . . . . . . . . . . . . . . . . . . .
8.3 Linear Regression and Other Linear Least Squares Problems
8.3.1 Example: Linear regression ...... .
8.3.2 Other least squares problems ...... .
8.4 Least Squares and Singular Value Decomposition
8.5 Least Squares and QR Factorization . . . . . . .
9 Eigenvalues and Eigenvectors
9.1 Fundamental Definitions and Properties
9.2 Jordan Canonical Form .... .
9.3 Determination of the JCF .... .
9.3.1 Theoretical computation .
9.3.2 On the + l's in JCF blocks
9.4 Geometric Aspects of the JCF
9.5 The Matrix Sign Function.
10 Canonical Forms
10.1 Some Basic Canonical Forms .
10.2 Definite Matrices . . . . . . .
10.3 Equivalence Transformations and Congruence
10.3.1 Block matrices and definiteness
10.4 Rational Canonical Form . . . . . . . . .
11 Linear Differential and Difference Equations
ILl Differential Equations . . . . . . . . . . . . . . . .
11.1.1 Properties ofthe matrix exponential . . . .
11.1.2 Homogeneous linear differential equations
11.1.3 Inhomogeneous linear differential equations
11.1.4 Linear matrix differential equations . .
11.1.5 Modal decompositions . . . . . . . . .
11.1.6 Computation of the matrix exponential
11.2 Difference Equations . . . . . . . . . . . . . .
11.2.1 Homogeneous linear difference equations
11.2.2 Inhomogeneous linear difference equations
11.2.3 Computation of matrix powers .
11.3 Higher-Order Equations. . . . . . . . . . . . . . .
Contents
51
51
52
54
57
59
65
65
67
67
67
69
70
71
75
75
82
85
86
88
89
91
95
95
99
102
104
104
109
109
109
112
112
113
114
114
118
118
118
119
120
Contents ix
12 Generalized Eigenvalue Problems 125
12.1 The Generalized Eigenvalue/Eigenvector Problem 125
12.2 Canonical Forms 127
12.3 Application to the Computation of System Zeros 130
12.4 Symmetric Generalized Eigenvalue Problems 131
12.5 Simultaneous Diagonalization 133
12.5.1 Simultaneous diagonalization via SVD 133
12.6 Higher-Order Eigenvalue Problems 135
12.6.1 Conversion to first-order form 135
13 Kronecker Products 139
13.1 Definition and Examples 139
13.2 Properties of the Kronecker Product 140
13.3 Application to Sylvester and Lyapunov Equations 144
Bibliography 151
Index 153
Contents
12 Generalized Eigenvalue Problems
12.1 The Generalized EigenvaluelEigenvector Problem
12.2 Canonical Forms ................ .
12.3 Application to the Computation of System Zeros .
12.4 Symmetric Generalized Eigenvalue Problems .
12.5 Simultaneous Diagonalization ........ .
12.5.1 Simultaneous diagonalization via SVD
12.6 Higher-Order Eigenvalue Problems ..
12.6.1 Conversion to first-order form
13 Kronecker Products
13.1 Definition and Examples ............ .
13.2 Properties of the Kronecker Product ...... .
13.3 Application to Sylvester and Lyapunov Equations
Bibliography
Index
ix
125
125
127
130
131
133
133
135
135
139
139
140
144
151
153
Preface
This book is intended to be used as a text for beginning graduate-level (or even senior-level)
students in engineering, the sciences, mathematics, computer science, or computational
science who wish to be familar with enough matrix analysis that they are prepared to use its
tools and ideas comfortably in a variety of applications. By matrix analysis I mean linear
algebra and matrix theory together with their intrinsic interaction with and application to
linear dynamical systems (systems of linear differential or difference equations). The text
can be used in a one-quarter or one-semester course to provide a compact overview of
much of the important and useful mathematics that, in many cases, students meant to learn
thoroughly as undergraduates, but somehow didn't quite manage to do. Certain topics
that may have been treated cursorily in undergraduate courses are treated in more depth
and more advanced material is introduced. I have tried throughout to emphasize only the
more important and "useful" tools, methods, and mathematical structures. Instructors are
encouraged to supplement the book with specific application examples from their own
particular subject area.
The choice of topics covered in linear algebra and matrix theory is motivated both by
applications and by computational utility and relevance. The concept of matrix factorization
is emphasized throughout to provide a foundation for a later course in numerical linear
algebra. Matrices are stressed more than abstract vector spaces, although Chapters 2 and 3
do cover some geometric (i.e., basis-free or subspace) aspects of many of the fundamental
notions. The books by Meyer [18], Noble and Daniel [20], Ortega [21], and Strang [24]
are excellent companion texts for this book. Upon completion of a course based on this
text, the student is then well-equipped to pursue, either via formal courses or through self-
study, follow-on topics on the computational side (at the level of [7], [11], [23], or [25], for
example) or on the theoretical side (at the level of [12], [13], or [16], for example).
Prerequisites for using this text are quite modest: essentially just an understanding
of calculus and definitely some previous exposure to matrices and linear algebra. Basic
concepts such as determinants, singularity of matrices, eigenvalues and eigenvectors, and
positive definite matrices should have been covered at least once, even though their recollec-
tion may occasionally be "hazy." However, requiring such material as prerequisite permits
the early (but "out-of-order" by conventional standards) introduction of topics such as pseu-
doinverses and the singular value decomposition (SVD). These powerful and versatile tools
can then be exploited to provide a unifying foundation upon which to base subsequent top-
ics. Because tools such as the SVD are not generally amenable to "hand computation," this
approach necessarily presupposes the availability of appropriate mathematical software on
a digital computer. For this, I highly recommend MA TL A B although other software such as
xi
Preface
This book is intended to be used as a text for beginning graduate-level (or even senior-level)
students in engineering, the sciences, mathematics, computer science, or computational
science who wish to be familar with enough matrix analysis that they are prepared to use its
tools and ideas comfortably in a variety of applications. By matrix analysis I mean linear
algebra and matrix theory together with their intrinsic interaction with and application to
linear dynamical systems (systems of linear differential or difference equations). The text
can be used in a one-quarter or one-semester course to provide a compact overview of
much of the important and useful mathematics that, in many cases, students meant to learn
thoroughly as undergraduates, but somehow didn't quite manage to do. Certain topics
that may have been treated cursorily in undergraduate courses are treated in more depth
and more advanced material is introduced. I have tried throughout to emphasize only the
more important and "useful" tools, methods, and mathematical structures. Instructors are
encouraged to supplement the book with specific application examples from their own
particular subject area.
The choice of topics covered in linear algebra and matrix theory is motivated both by
applications and by computational utility and relevance. The concept of matrix factorization
is emphasized throughout to provide a foundation for a later course in numerical linear
algebra. Matrices are stressed more than abstract vector spaces, although Chapters 2 and 3
do cover some geometric (i.e., basis-free or subspace) aspects of many of the fundamental
notions. The books by Meyer [18], Noble and Daniel [20], Ortega [21], and Strang [24]
are excellent companion texts for this book. Upon completion of a course based on this
text, the student is then well-equipped to pursue, either via formal courses or through self-
study, follow-on topics on the computational side (at the level of [7], [II], [23], or [25], for
example) or on the theoretical side (at the level of [12], [13], or [16], for example).
Prerequisites for using this text are quite modest: essentially just an understanding
of calculus and definitely some previous exposure to matrices and linear algebra. Basic
concepts such as determinants, singularity of matrices, eigenvalues and eigenvectors, and
positive definite matrices should have been covered at least once, even though their recollec-
tion may occasionally be "hazy." However, requiring such material as prerequisite permits
the early (but "out-of-order" by conventional standards) introduction of topics such as pseu-
doinverses and the singular value decomposition (SVD). These powerful and versatile tools
can then be exploited to provide a unifying foundation upon which to base subsequent top-
ics. Because tools such as the SVD are not generally amenable to "hand computation," this
approach necessarily presupposes the availability of appropriate mathematical software on
a digital computer. For this, I highly recommend MAlLAB although other software such as
xi
xii Preface
Mathematica or Mathcad is also excellent. Since this text is not intended for a course in
numerical linear algebra per se, the details of most of the numerical aspects of linear algebra
are deferred to such a course.
The presentation of the material in this book is strongly influenced by computa-
tional issues for two principal reasons. First, "real-life" problems seldom yield to simple
closed-form formulas or solutions. They must generally be solved computationally and
it is important to know which types of algorithms can be relied upon and which cannot.
Some of the key algorithms of numerical linear algebra, in particular, form the foundation
upon which rests virtually all of modern scientific and engineering computation. A second
motivation for a computational emphasis is that it provides many of the essential tools for
what I call "qualitative mathematics." For example, in an elementary linear algebra course,
a set of vectors is either linearly independent or it is not. This is an absolutely fundamental
concept. But in most engineering or scientific contexts we want to know more than that.
If a set of vectors is linearly independent, how "nearly dependent" are the vectors? If they
are linearly dependent, are there "best" linearly independent subsets? These turn out to
be much more difficult problems and frequently involve research-level questions when set
in the context of the finite-precision, finite-range floating-point arithmetic environment of
most modern computing platforms.
Some of the applications of matrix analysis mentioned briefly in this book derive
from the modern state-space approach to dynamical systems. State-space methods are
now standard in much of modern engineering where, for example, control systems with
large numbers of interacting inputs, outputs, and states often give rise to models of very
high order that must be analyzed, simulated, and evaluated. The "language" in which such
models are conveniently described involves vectors and matrices. It is thus crucial to acquire
a working knowledge of the vocabulary and grammar of this language. The tools of matrix
analysis are also applied on a daily basis to problems in biology, chemistry, econometrics,
physics, statistics, and a wide variety of other fields, and thus the text can serve a rather
diverse audience. Mastery of the material in this text should enable the student to read and
understand the modern language of matrices used throughout mathematics, science, and
engineering.
While prerequisites for this text are modest, and while most material is developed from
basic ideas in the book, the student does require a certain amount of what is conventionally
referred to as "mathematical maturity." Proofs are given for many theorems. When they are
not given explicitly, they are either obvious or easily found in the literature. This is ideal
material from which to learn a bit about mathematical proofs and the mathematical maturity
and insight gained thereby. It is my firm conviction that such maturity is neither encouraged
nor nurtured by relegating the mathematical aspects of applications (for example, linear
algebra for elementary state-space theory) to an appendix or introducing it "on-the-fly" when
necessary. Rather, one must lay a firm foundation upon which subsequent applications and
perspectives can be built in a logical, consistent, and coherent fashion.
I have taught this material for many years, many times at UCSB and twice at UC
Davis, and the course has proven to be remarkably successful at enabling students from
disparate backgrounds to acquire a quite acceptable level of mathematical maturity and
rigor for subsequent graduate studies in a variety of disciplines. Indeed, many students who
completed the course, especially the first few times it was offered, remarked afterward that
if only they had had this course before they took linear systems, or signal processing,
xii Preface
Mathematica or Mathcad is also excellent. Since this text is not intended for a course in
numerical linear algebra per se, the details of most of the numerical aspects of linear algebra
are deferred to such a course.
The presentation of the material in this book is strongly influenced by computa-
tional issues for two principal reasons. First, "real-life" problems seldom yield to simple
closed-form formulas or solutions. They must generally be solved computationally and
it is important to know which types of algorithms can be relied upon and which cannot.
Some of the key algorithms of numerical linear algebra, in particular, form the foundation
upon which rests virtually all of modem scientific and engineering computation. A second
motivation for a computational emphasis is that it provides many of the essential tools for
what I call "qualitative mathematics." For example, in an elementary linear algebra course,
a set of vectors is either linearly independent or it is not. This is an absolutely fundamental
concept. But in most engineering or scientific contexts we want to know more than that.
If a set of vectors is linearly independent, how "nearly dependent" are the vectors? If they
are linearly dependent, are there "best" linearly independent subsets? These tum out to
be much more difficult problems and frequently involve research-level questions when set
in the context of the finite-precision, finite-range floating-point arithmetic environment of
most modem computing platforms.
Some of the applications of matrix analysis mentioned briefly in this book derive
from the modem state-space approach to dynamical systems. State-space methods are
now standard in much of modem engineering where, for example, control systems with
large numbers of interacting inputs, outputs, and states often give rise to models of very
high order that must be analyzed, simulated, and evaluated. The "language" in which such
models are conveniently described involves vectors and matrices. It is thus crucial to acquire
a working knowledge of the vocabulary and grammar of this language. The tools of matrix
analysis are also applied on a daily basis to problems in biology, chemistry, econometrics,
physics, statistics, and a wide variety of other fields, and thus the text can serve a rather
diverse audience. Mastery of the material in this text should enable the student to read and
understand the modem language of matrices used throughout mathematics, science, and
engineering.
While prerequisites for this text are modest, and while most material is developed from
basic ideas in the book, the student does require a certain amount of what is conventionally
referred to as "mathematical maturity." Proofs are given for many theorems. When they are
not given explicitly, they are either obvious or easily found in the literature. This is ideal
material from which to learn a bit about mathematical proofs and the mathematical maturity
and insight gained thereby. It is my firm conviction that such maturity is neither encouraged
nor nurtured by relegating the mathematical aspects of applications (for example, linear
algebra for elementary state-space theory) to an appendix or introducing it "on-the-f1y" when
necessary. Rather, one must lay a firm foundation upon which subsequent applications and
perspectives can be built in a logical, consistent, and coherent fashion.
I have taught this material for many years, many times at UCSB and twice at UC
Davis, and the course has proven to be remarkably successful at enabling students from
disparate backgrounds to acquire a quite acceptable level of mathematical maturity and
rigor for subsequent graduate studies in a variety of disciplines. Indeed, many students who
completed the course, especially the first few times it was offered, remarked afterward that
if only they had had this course before they took linear systems, or signal processing.
Preface xiii
or estimation theory, etc., they would have been able to concentrate on the new ideas
they wanted to learn, rather than having to spend time making up for deficiencies in their
background in matrices and linear algebra. My fellow instructors, too, realized that by
requiring this course as a prerequisite, they no longer had to provide as much time for
"review" and could focus instead on the subject at hand. The concept seems to work.
AJL, June 2004
Preface XIII
or estimation theory, etc., they would have been able to concentrate on the new ideas
they wanted to learn, rather than having to spend time making up for deficiencies in their
background in matrices and linear algebra. My fellow instructors, too, realized that by
requiring this course as a prerequisite, they no longer had to provide as much time for
"review" and could focus instead on the subject at hand. The concept seems to work.
-AJL, June 2004
Chapter 1
Introduction and Review
We begin with a brief introduction to some standard notation and terminology to be used
throughout the text. This is followed by a review of some basic notions in matrix analysis
and linear algebra.
The following sets appear frequently throughout subsequent chapters:
1. R
n
= the set of n-tuples of real numbers represented as column vectors. Thus, x e Rn
means
where xi e R for i e n.
Henceforth, the notation n denotes the set {1, . . . , n}.
Note: Vectors are always column vectors. A row vector is denoted by y
T
, where
y G Rn and the superscript T is the transpose operation. That a vector is always a
column vector rather than a row vector is entirely arbitrary, but this convention makes
it easy to recognize immediately throughout the text that, e.g., X
T
y is a scalar while
xy
T
is an n x n matrix.
2. Cn = the set of n-tuples of complex numbers represented as column vectors.
3. R
mxn
= the set of real (or real-valued) m x n matrices.
4. R
mxnr
= the set of real m x n matrices of rank r. Thus, R
nxnn
denotes the set of real
nonsingular n x n matrices.
5. C
mxn
= the set of complex (or complex-valued) m x n matrices.
6. C
mxn
= the set of complex m x n matrices of rank r.
1
Chapter 1
Introduction and Review
We begin with a brief introduction to some standard notation and terminology to be used
throughout the text. This is followed by a review of some basic notions in matrix analysis
and linear algebra.
The following sets appear frequently throughout subsequent chapters:
I. IR
n
= the set of n-tuples of real numbers represented as column vectors. Thus, x E IR
n
means
where Xi E IR for i E !!.
Henceforth, the notation!! denotes the set {I, ... , n }.
Note: Vectors are always column vectors. A row vector is denoted by y ~ where
y E IR
n
and the superscript T is the transpose operation. That a vector is always a
column vector rather than a row vector is entirely arbitrary, but this convention makes
it easy to recognize immediately throughout the text that, e.g., x
T
y is a scalar while
xyT is an n x n matrix.
2. en = the set of n-tuples of complex numbers represented as column vectors.
3. IR
rn
xn = the set of real (or real-valued) m x n matrices.
4. 1R;n xn = the set of real m x n matrices of rank r. Thus, I R ~ xn denotes the set of real
nonsingular n x n matrices.
5. e
rnxn
= the set of complex (or complex-valued) m x n matrices.
6. e;n xn = the set of complex m x n matrices of rank r.
Chapter 1. Introduction and Review
Each of the above also has a "block" analogue obtained by replacing scalar components in
the respective definitions by block submatrices. For example, if A e R
nxn
, B e R
mx n
, and
C e R
mxm
, then the (m+ n) x (m+ n) matrix [ A0 Bc ] is block upper triangular.
The transpose of a matrix A is denoted by A
T
and is the matrix whose (i, j)th entry
is the (7, Oth entry of A, that is, (A
7
),, = a,,. Note that if A e R
mx
", then A
7
" e E"
xm
.
If A e C
mx
", then its Hermitian transpose (or conjugate transpose) is denoted by A
H
(or
sometimes A*) and its (i, j)\h entry is (A
H
),
7
= (77), where the bar indicates complex
conjugation; i.e., if z = a + jf$ (j = i = v^T), then z = a jfi. A matrix A is symmetric
if A = A
T
and Hermitian if A = A
H
. We henceforth adopt the convention that, unless
otherwise noted, an equation like A = A
T
implies that A is real-valued while a statement
like A = A
H
implies that A is complex-valued.
Remark 1.1. While \/\ is most commonly denoted by i in mathematics texts, j is
the more common notation in electrical engineering and system theory. There is some
advantage to being conversant with both notations. The notation j is used throughout the
text but reminders are placed at strategic locations.
Example 1.2.
Transposes of block matrices can be defined in an obvious way. For example, it is
easy to see that if A,, are appropriately dimensioned subblocks, then
is symmetric (and Hermitian).
is complex-valued symmetric but not Hermitian.
is Hermitian (but not symmetric).
2
We now classify some of the more familiar "shaped" matrices. A matrix A e
(or A eC"
x
")i s
diagonal if a,
7
= 0 for i ^ j.
upper triangular if a,
;
= 0 for i > j.
lower triangular if a,
7
= 0 for / < j.
tridiagonal if a
(y
= 0 for |z j\ > 1.
pentadiagonal if a
i;
= 0 for |/ j\ > 2.
upper Hessenberg if a
f
j = 0 for i j > 1.
lower Hessenberg if a,
;
= 0 for j i > 1.
2 Chapter 1. Introduction and Review
We now classify some of the more familiar "shaped" matrices. A matrix A E IR
n
xn
(or A E e
nxn
) is
diagonal if aij = 0 for i i= }.
upper triangular if aij = 0 for i > }.
lower triangular if aij = 0 for i < }.
tridiagonal if aij = 0 for Ii - JI > 1.
pentadiagonal if aij = 0 for Ii - J I > 2.
upper Hessenberg if aij = 0 for i - j > 1.
lower Hessenberg if aij = 0 for } - i > 1.
Each of the above also has a "block" analogue obtained by replacing scalar components in
the respective definitions by block submatrices. For example, if A E IR
nxn
, B E IR
nxm
, and
C E jRmxm, then the (m + n) x (m + n) matrix [ ~ ~ ] is block upper triangular.
The transpose of a matrix A is denoted by AT and is the matrix whose (i, j)th entry
is the (j, i)th entry of A, that is, (AT)ij = aji. Note that if A E jRmxn, then AT E jRnxm.
If A E em xn, then its Hermitian transpose (or conjugate transpose) is denoted by A H (or
sometimes A*) and its (i, j)th entry is (AH)ij = (aji), where the bar indicates complex
conjugation; i.e., if z = IX + jfJ (j = i = R), then z = IX - jfJ. A matrix A is symmetric
if A = A T and Hermitian if A = A H. We henceforth adopt the convention that, unless
otherwise noted, an equation like A = A T implies that A is real-valued while a statement
like A = AH implies that A is complex-valued.
Remark 1.1. While R is most commonly denoted by i in mathematics texts, } is
the more common notation in electrical engineering and system theory. There is some
advantage to being conversant with both notations. The notation j is used throughout the
text but reminders are placed at strategic locations.
Example 1.2.
1. A = [
; ~ ] is symmetric (and Hermitian).
2. A = [
5
7+}
7 + j ]
2 is complex-valued symmetric but not Hermitian.
[
5 7+} ]
3 A - 2 is Hermitian (but not symmetric).
- 7 - j
Transposes of block matrices can be defined in an obvious way. For example, it is
easy to see that if Aij are appropriately dimensioned subblocks, then
r = [
1.2. Matrix Arithmetic
1.2 Matrix Arithmetic
It is assumed that the reader is familiar with the fundamental notions of matrix addition,
multiplication of a matrix by a scalar, and multiplication of matrices.
A special case of matrix multiplication occurs when the second matrix is a column
vector x, i.e., the matrix-vector product Ax. A very important way to view this product is
to interpret it as a weighted sum (linear combination) of the columns of A. That is, suppose
The importance of this interpretation cannot be overemphasized. As a numerical example,
take A = [96 85 74]x = 2 . Then we can quickly calculate dot products of the rows of A
with the column x to find Ax =[50 32]' but this matrix-vector product can also be computed
v1a
For large arrays of numbers, there can be important computer-architecture-related advan-
tages to preferring the latter calculation method.
For matrix multiplication, suppose A e R
mxn
and B = [bi,...,b
p
] e R
nxp
with
bi e W
1
. Then the matrix product A B can be thought of as above, applied p times:
There is also an alternative, but equivalent, formulation of matrix multiplication that appears
frequently in the text and is presented below as a theorem. Again, its importance cannot be
overemphasized. It is deceptively simple and its full understanding is well rewarded.
Theorem 1.3. Let U = [M I , . . . , u
n
] e R
mxn
with u
t
e R
m
and V = [v
{
,..., v
n
] e R
pxn
with v
t
e R
p
. Then
If matrices C and D are compatible for multiplication, recall that (CD)
T
= D
T
C
T
(or (CD}
H
D
H
C
H
). This gives a dual to the matrix-vector result above. Namely, if
C eR
mxn
has row vectors cj e E
lx
", and is premultiplied by a row vector y
T
e R
l xm
,
then the product can be written as a weighted linear sum of the rows of C as follows:
3
Theorem 1.3 can then also be generalized to its "row dual." The details are left to the readei
Then
1.2. Matrix Arithmetic 3
1 .2 Matrix Arithmetic
It is assumed that the reader is familiar with the fundamental notions of matrix addition,
multiplication of a matrix by a scalar, and multiplication of matrices.
A special case of matrix multiplication occurs when the second matrix is a column
vector x, i.e., the matrix-vector product Ax. A very important way to view this product is
to interpret it as a weighted sum (linear combination) of the columns of A. That is, suppose
I ]
A = la' .... a"1 E JR
m
" with a, E JRm and x = l
Then
Ax = Xjal + ... + Xnan E jRm.
The importance of this interpretation cannot be overemphasized. As a numerical example,
take A = ! x = Then we can quickly calculate dot products of the rows of A
with the column x to find Ax = but this matrix-vector product can also be computed
via
3.[ J+2.[ J+l.[ l
For large arrays of numbers, there can be important computer-architecture-related advan-
tages to preferring the latter calculation method.
For matrix multiplication, suppose A E jRmxn and B = [hI,.'" h
p
] E jRnxp with
hi E jRn. Then the matrix product AB can be thought of as above, applied p times:
There is also an alternative, but equivalent, formulation of matrix multiplication that appears
frequently in the text and is presented below as a theorem. Again, its importance cannot be
overemphasized. It is deceptively simple and its full understanding is well rewarded.
Theorem 1.3. Let U = [Uj, ... , un] E jRmxn with Ui E jRm and V = [VI, .. , Vn] E lR
Pxn
with Vi E jRP. Then
n
UV
T
= LUiVr E jRmxp.
i=I
If matrices C and D are compatible for multiplication, recall that (C D)T = DT C
T
(or (C D)H = DH C
H
). This gives a dual to the matrix-vector result above. Namely, if
C E jRmxn has row vectors cJ E jRlxn, and is premultiplied by a row vector yT E jRlxm,
then the product can be written as a weighted linear sum of the rows of C as follows:
yTC=YICf EjRlxn.
Theorem 1.3 can then also be generalized to its "row dual." The details are left to the reader.
1.3 Inner Products and Orthogonality
For vectors x, y e R", the Euclidean inner product (or inner product, for short) of x and
y is given by
Note that the inner product is a scalar.
If x, y e C", we define their complex Euclidean inner product (or inner product,
for short) by
and we see that, indeed, (x, y)
c
= (y, x)
c
.
Note that x
T
x = 0 if and only if x = 0 when x e Rn but that this is not true if x e Cn.
What is true in the complex case is that X
H
x = 0 if and only if x = 0. To illustrate, consider
the nonzero vector x above. Then X
T
X = 0 but X
H
X = 2.
Two nonzero vectors x, y e R are said to be orthogonal if their inner product is
zero, i.e., x
T
y = 0. Nonzero complex vectors are orthogonal if X
H
y = 0. If x and y are
orthogonal and X
T
X = 1 and y
T
y = 1, then we say that x and y are orthonormal. A
matrix A e R
nxn
is an orthogonal matrix if A
T
A = AA
T
= /, where / is the n x n
identity matrix. The notation / is sometimes used to denote the identity matrix in R
nx
"
(orC"
x
"). Similarly, a matrix A e C
nxn
is said to be unitary if A
H
A = AA
H
= I. Clearly
an orthogonal or unitary matrix has orthonormal rows and orthonormal columns. There is
no special name attached to a nonsquare matrix A e R
mxn
(or C
mxn
) with orthonormal
rows or columns.
1.4 Determinants
It is assumed that the reader is familiar with the basic theory of determinants. For A e R
nxn
(or A 6 C
nxn
) we use the notation det A for the determinant of A. We list below some of
Note that (x, y)
c
= (y, x)
c
, i.e., the order in which x and y appear in the complex inner
product is important. The more conventional definition of the complex inner product is
( x , y )
c
= y
H
x = Eni=1 xiyi but throughout the text we prefer the symmetry with the real
case.
Example 1.4. Let x = [ 1j ] and y = [ 1/ 2 ]. Then
while
1.3 Inner Products and Orthogonality
For vectors x, y E IRn, the Euclidean inner product (or inner product, for short) of x and
y is given by
n
(x, y) := x
T
y = Lx;y;.
;=1
Note that the inner product is a scalar.
If x, y E <en, we define their complex Euclidean inner product (or inner product,
for short) by
n
(x'Y}c :=xHy = Lx;y;.
;=1
Note that (x, y)c = (y, x}c, i.e., the order in which x and y appear in the complex inner
product is important. The more conventional definition of the complex inner product is
(x, y)c = yH x = L:7=1 x;y; but throughout the text we prefer the symmetry with the real
case.
Example 1.4. Let x = [} ] and y = [ ~ ] . Then
(x, Y}c = [ } JH [ ] = [I - j] [ ~ ] = 1 - 2j
while
and we see that, indeed, (x, Y}c = {y, x)c'
Note that x
T
x = 0 if and only if x = 0 when x E IR
n
but that this is not true if x E en.
What is true in the complex case is that x
H
x = 0 if and only if x = O. To illustrate, consider
the nonzero vector x above. Then x
T
x = 0 but x
H
X = 2.
Two nonzero vectors x, y E IR
n
are said to be orthogonal if their inner product is
zero, i.e., x
T
y = O. Nonzero complex vectors are orthogonal if x
H
y = O. If x and y are
orthogonal and x
T
x = 1 and yT y = 1, then we say that x and y are orthonormal. A
matrix A E IR
nxn
is an orthogonal matrix if AT A = AAT = I, where I is the n x n
identity matrix. The notation In is sometimes used to denote the identity matrix in IR
nxn
(or en xn). Similarly, a matrix A E en xn is said to be unitary if A H A = AA H = I. Clearly
an orthogonal or unitary matrix has orthonormal rows and orthonormal columns. There is
no special name attached to a nonsquare matrix A E ]Rrn"n (or E e
mxn
) with orthonormal
rows or columns.
1.4 Determinants
It is assumed that the reader is familiar with the basic theory of determinants. For A E IR
n
xn
(or A E en xn) we use the notation det A for the determinant of A. We list below some of
1.4. Determinants
the more useful properties of determinants. Note that this is not a minimal set, i.e., several
properties are consequences of one or more of the others.
1. If A has a zero row or if any two rows of A are equal, then det A = 0.
2. If A has a zero column or if any two columns of A are equal, then det A = 0.
3. Interchanging two rows of A changes only the sign of the determinant.
4. Interchanging two columns of A changes only the sign of the determinant.
5. Multiplying a row of A by a scalar a results in a new matrix whose determinant is
a det A.
6. Multiplying a column of A by a scalar a results in a new matrix whose determinant
is a det A.
7. Multiplying a row of A by a scalar and then adding it to another row does not change
the determinant.
8. Multiplying a column of A by a scalar and then adding it to another column does not
change the determinant.
9. det A
T
= det A (det A
H
= det A if A e C
nxn
).
10. If A is diagonal, then det A = a11a22 a
nn
, i.e., det A is the product of its diagonal
elements.
11. If A is upper triangular, then det A = a11a22 a
nn
.
12. If A is lower triangular, then det A = a11a22 a
nn
.
13. If A is block diagonal (or block upper triangular or block lower triangular), with
square diagonal blocks A11, A22, , A
nn
(of possibly different sizes), then det A =
det A11 det A22 det A
nn
.
14. If A, B eR
nxn
,thendet(AB) = det A det 5.
15. If A R
nxn
, then det(A-
1
) = 1det A.
16. If A e R
nxn
and D e R
mxm
, then det [Ac
B
D
] = del A det ( D CA
l
B).
Proof: This follows easily from the block LU factorization
17. If A eR
nxn
and D e RM
mxm
, then det [Ac
B
D
] = det D det(A BD
1
C) .
Proof: This follows easily from the block UL factorization
5 1.4. Determinants 5
the more useful properties of determinants. Note that this is not a minimal set, i.e., several
properties are consequences of one or more of the others.
1. If A has a zero row or if any two rows of A are equal, then det A = o.
2. If A has a zero column or if any two columns of A are equal, then det A = O.
3. Interchanging two rows of A changes only the sign of the determinant.
4. Interchanging two columns of A changes only the sign of the determinant.
5. Multiplying a row of A by a scalar ex results in a new matrix whose determinant is
exdetA.
6. Multiplying a column of A by a scalar ex results in a new matrix whose determinant
is ex det A.
7. Multiplying a row of A by a scalar and then adding it to another row does not change
the determinant.
8. Multiplying a column of A by a scalar and then adding it to another column does not
change the determinant.
9. detAT = detA (detA
H
= detA if A E C"X").
10. If A is diagonal, then det A = alla22 ... ann, i.e., det A is the product of its diagonal
elements.
11. If A is upper triangular, then det A = all a22 ... a"n.
12. If A is lower triangUlar, then det A = alla22 ... ann.
13. If A is block diagonal (or block upper triangular or block lower triangular), with
square diagonal blocks A 11, A
22
, ... , An" (of possibly different sizes), then det A =
det A 11 det A22 ... det Ann.
14. If A, B E IR
nxn
, then det(AB) = det A det B.
15. If A E then det(A-
1
) = de: A .
16. If A E and DE IR
mxm
, then det = detA det(D - CA-
1
B).
Proof" This follows easily from the block LU factorization

] [
17. If A E IR
nxn
and D E then det = det D det(A - B D-
1
C).
Proof" This follows easily from the block UL factorization
BD-
1
I
] [
Remark 1.5. The factorization of a matrix A into the product of a unit lower triangular
matrix L (i.e., lower triangular with all 1's on the diagonal) and an upper triangular matrix
U is called an LU factorization; see, for example, [24]. Another such factorization is UL
where U is unit upper triangular and L is lower triangular. The factorizations used above
are block analogues of these.
Remark 1.6. The matrix D CA
1
B is called the Schur complement of A in [AC BD].
Similarly, A BD
l
C is the Schur complement of D in [AC
B
D
].
EXERCISES
1. If A e R
nxn
and or is a scalar, what is det(aA)? What is det(A)?
2. If A is orthogonal, what is det A? If A is unitary, what is det A?
3. Let x, y e Rn. Show that det(I xy
T
) = 1 y
T
x.
4. Let U1, U
2
, . . ., Uk R
nxn
be orthogonal matrices. Show that the product U =
U1 U2 Uk is an orthogonal matrix.
5. Let A e R
n x n
. The trace of A, denoted TrA, is defined as the sum of its diagonal
elements, i.e., TrA = Eni=1
aii.
(a) Show that the trace is a linear function; i.e., if A, B e R
nxn
and a, ft e R, then
Tr(aA + fiB)= aTrA + fiTrB.
(b) Show that Tr(Afl) = Tr(A), even though in general AB ^ B A.
(c) Let S R
nxn
be skew-symmetric, i.e., S
T
= -S. Show that TrS = 0. Then
either prove the converse or provide a counterexample.
6. A matrix A e W
x
" is said to be idempotent if A
2
= A.
/ x . , , ! T 2cos
2
<9 sin 20 1 . . _ ,
(a) Show that the matrix A = - . _ .. _ .
2rt
is idempotent for all #.
2 |_ sin 2^ 2sm
z
# J
r
(b) Suppose A e IR"
X
" is idempotent and A ^ I. Show that A must be singular.
Remark 1.5. The factorization of a matrix A into the product of a unit lower triangular
matrix L (i.e., lower triangular with all l's on the diagonal) and an upper triangular matrix
V is called an LV factorization; see, for example, [24]. Another such factorization is VL
where V is unit upper triangular and L is lower triangular. The factorizations used above
are block analogues of these.
Remark 1.6. The matrix D - e A -I B is called the Schur complement of A in [ ~ ~ ].
Similarly, A - BD-Ie is the Schur complement of Din ~ ~ l
EXERCISES
1. If A E jRnxn and a is a scalar, what is det(aA)? What is det(-A)?
2. If A is orthogonal, what is det A? If A is unitary, what is det A?
3. Letx,y E jRn. Showthatdet(l-xyT) = 1- yTx.
4. Let VI, V2, ... ,Vk E jRn xn be orthogonal matrices. Show that the product V =
VI V2 ... V
k
is an orthogonal matrix.
5. Let A E jRNxn. The trace of A, denoted Tr A, is defined as the sum of its diagonal
elements, i.e., TrA = L ~ = I au
(a) Show that the trace is a linear function; i.e., if A, B E JRn xn and a, f3 E JR, then
Tr(aA + f3B) = aTrA + f3TrB.
(b) Show that Tr(AB) = Tr(BA), even though in general AB i= BA.
(c) Let S E jRnxn be skew-symmetric, i.e., ST = -So Show that TrS = O. Then
either prove the converse or provide a counterexample.
6. A matrix A E jRnxn is said to be idempotent if A2 = A.
I [ 2cos
2
0
(a) Show that the matrix A = - . 2f)
2 sm 0
sin 20 J. . d .. II II
2sin
2
0 IS I empotent lor a o.
(b) Suppose A E jRn xn is idempotent and A i= I. Show that A must be singular.
Chapter 2
Vector Spaces
In this chapter we give a brief review of some of the basic concepts of vector spaces. The
emphasis is on finite-dimensional vector spaces, including spaces formed by special classes
of matrices, but some infinite-dimensional examples are also cited. An excellent reference
for this and the next chapter is [10], where some of the proofs that are not given here may
be found.
2.1 Definitions and Examples
Definition 2.1. A field is a set F together with two operations +, : F x F > F such that
Axioms (A1)-(A3) state that (F, +) is a group and an abelian group if (A4) also holds.
Axioms (M1)-(M4) state that (F \ {0}, ) is an abelian group.
Generally speaking, when no confusion can arise, the multiplication operator "" is
not written explicitly.
7
(Al) a + (P + y ) = (a + p ) + y f o r all a, f t, y F.
(A2) there exists an element 0 e F such that a + 0 = a. for all a e F.
(A3 ) for all a e F, there exists an element (a) e F such that a + ( a) = 0.
(A4 ) a + p = ft + afar all a, ft e F.
(M l) a - ( p - y ) = ( a - p ) - y f o r al l a, p, y e F.
(M 2) there exists an element 1 e F such that a I = a for all a e F.
(M 3 ) for all a e , a ^0, there exists an element a"
1
F such that a a~
l
= 1.
(M 4 ) a p = P a for all a, p e F.
(D) a - ( p + y)=ci- p+a- y f or alia, p,ye.
Chapter 2
Vector Spaces
In this chapter we give a brief review of some of the basic concepts of vector spaces. The
emphasis is on finite-dimensional vector spaces, including spaces formed by special classes
of matrices, but some infinite-dimensional examples are also cited. An excellent reference
for this and the next chapter is [10], where some of the proofs that are not given here may
be found.
2.1 Definitions and Examples
Definition 2.1. A field is a set IF together with two operations +, . : IF x IF ~ IF such that
(Al) a + (,8 + y) = (a +,8) + y for all a,,8, y Elf.
(A2) there exists an element 0 E IF such that a + 0 = a for all a E IF.
(A3) for all a E IF, there exists an element (-a) E IF such that a + (-a) = O.
(A4) a + ,8 = ,8 + a for all a, ,8 Elf.
(Ml) a (,8, y) = (a,8) y for all a,,8, y Elf.
(M2) there exists an element I E IF such that a . I = a for all a E IF.
(M3) for all a E IF, a f. 0, there exists an element a-I E IF such that a . a-I = 1.
(M4) a,8 =,8 . afar all a, ,8 E IF.
(D) a (,8 + y) = a,8 +a y for all a, ,8, y Elf.
Axioms (Al)-(A3) state that (IF, +) is a group and an abelian group if (A4) also holds.
Axioms (MI)-(M4) state that (IF \ to), .) is an abelian group.
Generally speaking, when no confusion can arise, the multiplication operator "." is
not written explicitly.
7
Chapter 2. Vector Spaces
Example 2.2.
1. R with ordinary addition and multiplication is a field.
2. C with ordinary complex addition and multiplication is a field.
3. Raf. r] = the field of rational functions in the indeterminate x
8
where Z+ = {0,1,2, . . . }, is a field.
4. RMr
mxn
= { m x n matrices of rank r with real coefficients) is clearly not a field since,
for example, (Ml) does not hold unless m = n. Moreover, R"
x
" is not a field either
since (M4) does not hold in general (although the other 8 axioms hold).
Definition 2.3. A vector space over a field F is a set V together with two operations
+ :V x V -^V and- : F xV - V such that
A vector space is denoted by (V, F) or, when there is no possibility of confusion as to the
underlying fie Id, simply by V.
Remark 2.4. Note that + and in Definition 2.3 are different from the + and in Definition
2.1 in the sense of operating on different objects in different sets. In practice, this causes
no confusion and the operator is usually not even written explicitly.
Example 2.5.
1. (R", R) with addition defined by
and scalar multiplication defined by
is a vector space. Similar definitions hold for (C", C).
(VI) (V, +) is an abelian group.
(V2) ( a - p ) - v = a - ( P ' V ) f o r all a, p e F and for all v e V.
(V3) (a + ft) v = a v + p v for all a, p F and for all v e V.
(V4) a-(v + w)=a-v + a- w for all a e F and for all v, w e V.
(V5) 1 v = v for all v e V (1 e F).
8 Chapter 2. Vector Spaces
Example 2.2.
I. IR with ordinary addition and multiplication is a field.
2. e with ordinary complex addition and multiplication is a field.
3. Ra[x] = the field of rational functions in the indeterminate x
= {a
o
+ atX + ... + apxP +}
:aj,f3i EIR ;P,qEZ ,
f30 + f3t
X
+ ... + f3qX
q
where Z+ = {O,l,2, ... }, is a field.
4. I R ~ xn = { m x n matrices of rank r with real coefficients} is clearly not a field since,
for example, (MI) does not hold unless m = n. Moreover, l R ~ x n is not a field either
since (M4) does not hold in general (although the other 8 axioms hold).
Definition 2.3. A vector space over a field IF is a set V together with two operations
+ : V x V -+ V and : IF x V -+ V such that
(VI) (V, +) is an abelian group.
(V2) (a f3) . v = a . (f3 . v) for all a, f3 E IF andfor all v E V.
(V3) (a + f3). v = a v + f3. v for all a, f3 Elf andforall v E V.
(V4) a (v + w) = a . v + a . w for all a ElF andfor all v, w E V.
(V5) I v = v for all v E V (1 Elf).
A vector space is denoted by (V, IF) or, when there is no possibility of confusion as to the
underlying field, simply by V.
Remark 2.4. Note that + and in Definition 2.3 are different from the + and . in Definition
2.1 in the sense of operating on different objects in different sets. In practice, this causes
no confusion and the operator is usually not even written explicitly.
Example 2.5.
I. (IRn, IR) with addition defined by
is a vector space. Similar definitions hold for (en, e).
2.2. Subspaces
3. Let (V, F) be an arbitrary vector space and V be an arbitrary set. Let O (X > , V) be the
set of functions / mapping D to V. Then O (D, V) is a vector space with addition
defined by
2.2 Subspaces
Definition 2.6. Let (V, F) be a vector space and let W c V, W = 0. Then (W, F) is a
subspace of (V, F) i f and only i f (W, F) is i tself a vector space or, equi valently, i f and only
i f ( a w 1 + W 2 ) e W for all a, e and for all w 1 , w
2
e W.
Remark 2.7. The latter characterization of a subspace is often the easiest way to check
or prove that something is indeed a subspace (or vector space); i.e., verify that the set in
question is closed under addition and scalar multiplication. Note, too, that since 0 e F, this
implies that the zero vector must be in any subspace.
Notation: When the underlying field is understood, we write W c V, and the symbol c,
when used with vector spaces, is henceforth understood to mean "is a subspace of." The
less restrictive meaning "is a subset of" is specifically flagged as such.
9
2. (E
mxn
, E) is a vector space with addition defined by
Special Cases:
(a) V = [to, t \ ] , (V, F) = (IR", E), and the functions are piecewise continuous
=: (PC[f
0
, t\ ] )
n
or continuous =: (C[?
0
, h] )
n
.
4. Let A R"
x
". Then (x(t) : x ( t ) = Ax(t}} is a vector space (of dimension n) .
2.2. Subspaces 9
2.
(JRmxn, JR) is a vector space with addition defined by
[ ." + P"
al2 + fJI2 aln + fJln
l
a21 + fJ2I a22 + fJ22 a2n + fJ2n
A+B= .
amI + fJml am2 + fJm2 amn + fJmn
[ ya"
y
a
l2
ya," l
y
a
21 y
a
22 ya2n
yA = . . .
yaml ya
m
2
ya
mn
3. Let (V, IF) be an arbitrary vector space and '0 be an arbitrary set. Let cf>('O, V) be the
set of functions f mapping '0 to V. Then cf>('O, V) is a vector space with addition
defined by
(f + g)(d) = fed) + g(d) for all d E '0 and for all f, g E cf>
(af)(d) = af(d) for all a E IF, for all d ED, and for all f E cf>.
Special Cases:
(a) '0 = [to, td, (V, IF) = (JR
n
, JR), and the functions are piecewise continuous
=: (PC[to, td)n or continuous =: (C[to, td)n.
(b) '0 = [to, +00), (V, IF) = (JRn, JR), etc.
4. Let A E JR(nxn. Then {x(t) : x(t) = Ax(t)} is a vector space (of dimension n).
2.2 Subspaces
Definition 2.6. Let (V, IF) be a vector space and let W ~ V, W f= 0. Then (W, IF) is a
subspace of (V, IF) if and only if (W, IF) is itself a vector space or, equivalently, if and only
if(awl + fJw2) E W foral! a, fJ E IF andforall WI, W2 E W.
Remark 2.7. The latter characterization of a subspace is often the easiest way to check
or prove that something is indeed a subspace (or vector space); i.e., verify that the set in
question is closed under addition and scalar multiplication. Note, too, that since 0 E IF, this
implies that the zero vector must be in any subspace.
Notation: When the underlying field is understood, we write W ~ V, and the symbol ~ ,
when used with vector spaces, is henceforth understood to mean "is a subspace of." The
less restrictive meaning "is a subset of' is specifically flagged as such.
Then W
a
, is a subspace of V if and only if = 0. As an interesting exercise, sketch
W2,1, W2,o, W1/2,1, and W1/2,
0
. Note, too, that the vertical line through the origin (i.e.,
a = oo) is also a subspace.
All lines through the origin are subspaces. Shifted subspaces W
a
, with = 0 are
called linear varieties.
Henceforth, we drop the explicit dependence of a vector space on an underlying field.
Thus, V usually denotes a vector space with the underlying field generally being R unless
explicitly stated otherwise.
Definition 2.9. If 12, and S are vector spaces (or subspaces), then R = S if and only if
R C S and S C R.
Note: To prove two vector spaces are equal, one usually proves the two inclusions separately:
An arbitrary r e R is shown to be an element of S and then an arbitrary 5 S is shown to
be an element of R.
2.3 Linear Independence
Let X = { v1 , v2, } be a nonempty collection of vectors u, in some vector space V.
Definition 2.10. X is a linearly dependent set of vectors if and only if there exist k distinct
elements v1, . . . , vk e X and scalars a1, . . . , ak not all zero such that
Example 2.8.
1. Consider (V, F) = (R"
X
",R) and let W = [A e R"
x
" : A is symmetric}. Then
We V.
Proof: Suppose A\, A
2
are symmetric. Then it is easily shown that ctA\ + fiAi is
symmetric for all a, ft e R.
2. Let W = { A R"
x
" : A is orthogonal}. Then W is /wf a subspace of R"
x
".
3. Consider (V, F) = (R
2
, R) and for each v R
2
of the form v = [v1v2 ] identify v1 with
the jc-coordinate in the plane and u
2
with the y-coordinate. For a, e R, define
X is a linearly independent set of vectors if and only if for any collection of k distinct
elements v1, . . . ,Vk of X and for any scalars a1, . . . , ak,
Example 2.S.
1. Consider (V,lF) = (JR.nxn,JR.) and let W = {A E JR.nxn : A is symmetric}. Then

Proof' Suppose AI, A2 are symmetric. Then it is easily shown that aAI + f3A2 is
symmetric for all a, f3 E R
2. Let W = {A E ]Rnxn : A is orthogonal}. Then W is not a subspace of JR.nxn.
3. Consider (V, IF) = (]R2, JR.) and for each v E ]R2 of the form v = ] identify VI with
the x-coordinate in the plane and V2 with the y-coordinate. For a, f3 E R define
W",/l = {V : v = [ ac f3 ] ; c E JR.} .
Then W",/l is a subspace of V if and only if f3 = O. As an interesting exercise, sketch
W2.I, W2,O, Wi,I' and Wi,o, Note, too, that the vertical line through the origin (i.e.,
a = 00) is also a subspace.
All lines through the origin are subspaces. Shifted subspaces W",/l with f3 =1= 0 are
called linear varieties.
Henceforth, we drop the explicit dependence of a vector space on an underlying field.
Thus, V usually denotes a vector space with the underlying field generally being JR. unless
explicitly stated otherwise.
Definition 2.9. ffR and S are vector spaces (or subspaces), then R = S if and only if
R R.
Note: To prove two vector spaces are equal, one usually proves the two inclusions separately:
An arbitrary r E R is shown to be an element of S and then an arbitrary s E S is shown to
be an element of R.
2.3 Linear Independence
Let X = {VI, V2, . } be a nonempty collection of vectors Vi in some vector space V.
Definition 2.10. X is a linearly dependent set of vectors if and only if there exist k distinct
elements VI, ... , Vk E X and scalars aI, .. , (Xk not all zero such that
X is a linearly independent set of vectors if and only if for any collection of k distinct
elements VI, ... , Vk of X and for any scalars aI, , ak,
al VI + ... + (XkVk = 0 implies al = 0, ... , ak = O.
2.3. Linear Independence 11
(since 2v\ v
2
+ v3 = 0).
2. Let A e R
xn
and 5 e R"
xm
. Then consider the rows of e
tA
B as vectors in C
m
[t
0
, t1]
(recall that e
fA
denotes the matrix exponential, which is discussed in more detail in
Chapter 11). Independence of these vectors turns out to be equivalent to a concept
called controllability, to be studied further in what follows.
Let v
f
e R", i e k, and consider the matrix V = [ v1 , ... ,Vk] e R
nxk
. The linear
dependence of this set of vectors is equivalent to the existence of a nonzero vector a e R
k
such that Va = 0. An equivalent condition for linear dependence is that the k x k matrix
V
T
V is singular. If the set of vectors is independent, and there exists a e R* such that
Va = 0, then a = 0. An equivalent condition for linear independence is that the matrix
V
T
V is nonsingular.
Definition 2.12. Let X = [ v1 , v2, . . . } be a collection of vectors vi. e V. Then the span of
X is defined as
Example 2.13. Let V = R
n
and define
Then Sp{e1, e
2
, ...,e
n
} = Rn.
Definition 2.14. A set of vectors X is a basis for V if and only ij
1. X is a linearly independent set (of basis vectors), and
2. Sp(X) = V.
Example 2.11.
is a linearly independent set. Why?
s a linearly dependent set However,
1. LetV = R
3
. Then
where N = {1, 2, ...}.
2.3. Linear Independence 11
Example 2.11.
I. 1t V = Then {[ H i Hi] } i" independent.. Why?
Howe,."I [ i 1 [ i 1 [ l ] } is a Iin=ly
(since 2vI - V2 + V3 = 0).
2. Let A E ]Rnxn and B E ]Rnxm. Then consider the rows of etA B as vectors in em [to, tIl
(recall that etA denotes the matrix exponential, which is discussed in more detail in
Chapter 11). Independence of these vectors turns out to be equivalent to a concept
called controllability, to be studied further in what follows.
Let Vi E ]Rn, i E If, and consider the matrix V = [VI, ... , Vk] E ]Rnxk. The linear
dependence of this set of vectors is equivalent to the existence of a nonzero vector a E ]Rk
such that Va = O. An equivalent condition for linear dependence is that the k x k matrix
VT V is singular. If the set of vectors is independent, and there exists a E ]Rk such that
Va = 0, then a = O. An equivalent condition for linear independence is that the matrix
V T V is nonsingular.
Definition 2.12. Let X = {VI, V2, .. } be a collection of vectors Vi E V. Then the span of
X is defined as
Sp(X) = Sp{VI, V2, ... }
= {v : V = (Xl VI + ... + (XkVk ; (Xi ElF, Vi EX, kEN},
where N = {I, 2, ... }.
Example 2.13. Let V = ]Rn and define
0 0
0 1 0
el =
0
, e2 =
0
,'" ,en =
0
o o
Then SpIel, e2, ... , en} = ]Rn.
Definition 2.14. A set of vectors X is a basis for V if and only if
1. X is a linearly independent set (of basis vectors), and
2. Sp(X) = V.
Example 2.15. [e\,..., e
n
} is a basis for IR" (sometimes called the natural basis).
Now let b1, ..., b
n
be a basis (with a specific order associated with the basis vectors)
for V. Then for all v e V there exists a unique n-tuple {E1 , . . . , E n} such that
Definition 2.16. The scalars {Ei} are called the components (or sometimes the coordinates)
of v with respect to the basis (b1, ..., b
n
] and are unique. We say that the vector x of
components represents the vector v with respect to the basis B.
Example 2.17. In Rn,
we have
To see this, write
Then
Theorem 2.18. The number of elements in a basis of a vector space is independent of the
particular basis considered.
Definition 2.19. If a basis X for a vector space V= 0) has n elements, V is said to
be n-dimensional or have dimension n and we write dim(V) = n or dim V n. For
We can also determine components of v with respect to another basis. For example, while
with respect to the basis
where
Example 2.15. {el, ... , en} is a basis for]Rn (sometimes called the natural basis).
Now let b
l
, ... , b
n
be a basis (with a specific order associated with the basis vectors)
for V. Then for all v E V there exists a unique n-tuple ... , such that
v = + ... + = Bx,
where
B [b".,b.l. x D J
Definition 2.16. The scalars } are called the components (or sometimes the coordinates)
of v with respect to the basis {b
l
, ... , b
n
} and are unique. We say that the vector x of
components represents the vector v with respect to the basis B.
Example 2.17. In]Rn,
VI ]
: = vlel + V2e2 + ... + vne
n

Vn
We can also determine components of v with respect to another basis. For example, while
with respect to the basis
we have
To see this, write
Then
[ ] = I . el + 2 . e2,

[ ] = 3 . [ ] + 4 [ l
[ ] = XI [ - ] + X2 [ _! ]
= [ - -! ] [ l
[ ] = [ -; -1 r I [ ; ] = [ l
Theorem 2.18. The number of elements in a basis of a vector space is independent of the
particular basis considered.
Definition 2.19. If a basis X for a vector space V(Jf 0) has n elements, V is said to
be n.dimensional or have dimension n and we write dim (V) = n or dim V = n. For
Definition 2.21. Let (V, F) be a vector space and let 71, S c V. The sum and intersection
of R, and S are defined respectively by:
The subspaces R, and S are said to be complements of each other in T.
Remark 2.23. The union of two subspaces, R C S, is not necessarily a subspace.
Definition 2.24. T = R 0 S is the direct sum of R and S if
Theorem 2.22.
2.4. Sums and Intersections of Subspaces 13
consistency, and because the 0 vector is in any vector space, we define dim(O) = 0. A
vector space V is finite-dimensional if there exists a basis X with n < +00 elements;
otherwise, V is infinite-dimensional.
Thus, Theorem 2.18 says that dim(V) = the number of elements in a basis.
Example 2.20.
1. d i m(Rn)=n.
2. dim(R
mxn
) = mn.
Note: Check that a basis for R
mxn
is given by the mn matrices Eij; i e m, j e n,
where E
f
j is a matrix all of whose elements are 0 except for a 1 in the (i, j)th location.
The collection of Eij matrices can be called the "natural basis matrices."
3. dim(C[to, t1]) - +00.
4. dim{A R
nxn
: A = A
T
} = {1/2(n + 1).
1
2
(To see why, determine 1/ 2n( n + 1) symmetric basis matrices.)
5. dim{A e R
nxn
: A is upper (lower) triangular} = 1/ 2n( n + 1).
1. n + S = {r + s : r e U, s e 5}.
2. ft H 5 = {v : v e 7^ and v e 5}.
K
1. K + S C V (in general, U\ - \ h 7^ =: ]T ft/ C V, for finite k).
1=1
2. 72. D 5 C V (in general, f] * R,
a
C V/ or an arbitrary index set A).
a e A
1. n n S = 0, and
2. U + S = T (in general ft; n (^ ft,-) = 0 am/ ]Pft,- = T).
y>f
2.4. Sums and Intersections of Subspaces 13
consistency, and because the 0 vector is in any vector space, we define dim(O) = O. A
vector space V is finite-dimensional if there exists a basis X with n < +00 elements;
otherwise, V is infinite-dimensional.
Thus, Theorem 2.18 says that dim (V) = the number of elements in a basis.
Example 2.20.
1. = n.
2. = mn.
Note: Check that a basis for is given by the mn matrices Eij; i E m, j E
where Eij is a matrix all of whose elements are 0 except for a 1 in the (i, J)th location.
The collection of E;j matrices can be called the "natural basis matrices."
3. dim(C[to, tJJ) = +00.
4. dim{A E : A = AT} = !n(n + 1).
(To see why, determine !n(n + 1) symmetric basis matrices.)
5. dim{A E : A is upper (lower) triangular} = !n(n + 1).
Definition 2.21. Let (V, JF') be a vector space and let R, S S; V. The sum and intersection
ofR and S are defined respectively by:
1. R + S = {r + s : r E R, s E S}.
2. R n S = {v : v E R and v E S}.
Theorem 2.22.
k
1. R + S S; V (in general, RI + ... + Rk =: L R; S; V, for finite k).
;=1
2. R n S S; V (in general, n Ra S; V for an arbitrary index set A).
CiEA
Remark 2.23. The union of two subspaces, R U S, is not necessarily a subspace.
Definition 2.24. T = REB S is the direct sum ofR and S if
1. R n S = 0, and
2. R + S = T (in general, R; n (L R
j
) = 0 and L Ri = T).
H;
The subspaces Rand S are said to be complements of each other in T.
Remark 2.25. The complement of ft (or S) is not unique. For example, consider V = R
2
and let ft be any line through the origin. Then any other distinct line through the origin is
a complement of ft. Among all the complements there is a unique one orthogonal to ft.
We discuss more about orthogonal complements elsewhere in the text.
Theorem 2.26. Suppose T =R O S. Then
1. every t T can be written uniquely in the form t = r + s with r e R and s e S.
2. dim(T) = dim(ft) + dim(S).
Proof: To prove the first part, suppose an arbitrary vector t e T can be written in two ways
as t = r1 + s1 = r2 + S2, where r1, r2 e R. and s1, S2 e S. Then r1 r2 = s2 s\. But
r1 r2 ft and 52 si e S. Since ft fl S = 0, we must have r\ = r-i and s\ = si from
which uniqueness follows.
The statement of the second part is a special case of the next theorem. D
Theorem 2.27. For arbitrary subspaces ft, S of a vector space V,
EXERCISES
1. Suppose {vi,..., Vk} is a linearly dependent set. Then show that one of the vectors
must be a linear combination of the others.
2. Let x\, *2, . . . , x/c E R" be nonzero mutually orthogonal vectors. Show that [x\,...,
X k} must be a linearly independent set.
3. Let v\,... ,v
n
be orthonormal vectors in R". Show that Av\,..., Av
n
are also or-
thonormal if and only if Ae R"
x
" is orthogonal.
4. Consider the vectors v\ [2 l]
r
and 1*2 = [3 l]
r
. Prove that vi and V2 form a basis
for R
2
. Find the components of the vector v = [4 l]
r
with respect to this basis.
Example 2.28. Let U be the subspace of upper triangular matrices in E"
x
" and let be the
subspace of lower triangular matrices in R
nxn
. Then it may be checked that U + L = R
nxn
while U n is the set of diagonal matrices in R
nxn
. Using the fact that dim (diagonal
matrices} = n, together with Examples 2.20.2 and 2.20.5, one can easily verify the validity
of the formula given in Theorem 2.27.
Example 2.29. Let (V, F) = (R
nxn
, R), let ft be the set of skew-symmetric matrices in
R"
x
", and let S be the set of symmetric matrices in R"
x
". Then V = U 0 S.
Proof: This follows easily from the fact that any Ae R"
x
" can be written in the form
The first matrix on the right-hand side above is in S while the second is in ft.
Remark 2.25. The complement of R (or S) is not unique. For example, consider V = jR2
and let R be any line through the origin. Then any other distinct line through the origin is
a complement of R. Among all the complements there is a unique one orthogonal to R.
We discuss more about orthogonal complements elsewhere in the text.
Theorem 2.26. Suppose T = R EB S. Then
1. every t E T can be written uniquely in the form t = r + s with r E Rand s E S.
2. dim(T) = dim(R) + dim(S).
Proof: To prove the first part, suppose an arbitrary vector t E T can be written in two ways
as t = rl + Sl = r2 + S2, where rl, r2 E Rand SI, S2 E S. Then r, - r2 = S2 - SI. But
rl - r2 E Rand S2 - SI E S. Since R n S = 0, we must have rl = r2 and SI = S2 from
which uniqueness follows.
The statement of the second part is a special case of the next theorem. 0
Theorem 2.27. For arbitrary subspaces R, S of a vector space V,
dim(R + S) = dim(R) + dim(S) - dim(R n S).
Example 2.28. Let U be the subspace of upper triangular matrices in jRn xn and let .c be the
subspace of lower triangUlar matrices in jRn xn. Then it may be checked that U + .c = jRn xn
while un.c is the set of diagonal matrices in jRnxn. Using the fact that dim {diagonal
matrices} = n, together with Examples 2.20.2 and 2.20.5, one can easily verify the validity
of the formula given in Theorem 2.27.
Example 2.29. Let (V, IF) = (jRnxn, jR), let R be the set of skew-symmetric matrices in
jRnxn, and let S be the set of symmetric matrices in jRnxn. Then V = n $ S.
Proof: This follows easily from the fact that any A E jRnxn can be written in the form
1 TIT
A=2:(A+A )+2:(A-A).
The first matrix on the right-hand side above is in S while the second is in R.
EXERCISES
1. Suppose {VI, ... , vd is a linearly dependent set. Then show that one of the vectors
must be a linear combination of the others.
2. Let XI, X2, ... , Xk E jRn be nonzero mutually orthogonal vectors. Show that {XI, ... ,
Xk} must be a linearly independent set.
3. Let VI, ... , Vn be orthonormal vectors in jRn. Show that Av" .. , AV
n
are also or-
thonormal if and only if A E jRnxn is orthogonal.
4. Consider the vectors VI = [2 1 f and V2 = [3 1 f. Prove that VI and V2 form a basis
for jR2. Find the components of the vector v = [4 If with respect to this basis.
Exercises 15
5. Let P denote the set of polynomials of degree less than or equal to two of the form
Po + p\x + pix
2
, where po, p\, p2 e R. Show that P is a vector space over E. Show
that the polynomials 1, *, and 2x
2
1 are a basis for P. Find the components of the
polynomial 2 + 3x + 4x
2
6. Prove Theorem 2.22 (for the case of two subspaces R and S only).
7. Let P
n
denote the vector space of polynomials of degree less than or equal to n, and of
the form p( x) = po + p\x + + p
n
x
n
, where the coefficients /?, are all real. Let PE
denote the subspace of all even polynomials in P
n
, i.e., those that satisfy the property
p(x} = p(x). Similarly, let PQ denote the subspace of all odd polynomials, i.e.,
those satisfying p(x} = p( x) . Show that P
n
= P
E
PO-
8. Repeat Example 2.28 using instead the two subspaces 7" of tridiagonal matrices and
U of upper triangular matrices.
Exercises 15
5. Let P denote the set of polynomials of degree less than or equal to two of the form
Po + PI X + P2x2, where Po, PI, P2 E R Show that P is a vector space over R Show
that the polynomials 1, x, and 2x2 - 1 are a basis for P. Find the components of the
polynomial 2 + 3x + 4x
2
6. Prove Theorem 2.22 (for the case of two subspaces Rand S only).
7. Let P
n
denote the vector space of polynomials of degree less than or equal to n, and of
the form p(x) = Po + PIX + ... + Pnxn, where the coefficients Pi are all real. Let PE
denote the subspace of all even polynomials in P
n
, i.e., those that satisfy the property
p( -x) = p(x). Similarly, let Po denote the subspace of all odd polynomials, i.e.,
those satisfying p(-x) = -p(x). Show that P
n
= PE EB Po
8. Repeat Example 2.28 using instead the two subspaces T of tridiagonal matrices and
U of upper triangular matrices.
Chapter 3
Linear Transformations
3.1 Definition and Examples
We begin with the basic definition of a linear transformation (or linear map, linear function,
or linear operator) between two vector spaces.
Definition 3.1. Let (V, F) and (W, F) be vector spaces. Then C : V -> W is a linear
transformation if and only if
(avi + pv
2
) = aCv\ + fiv
2
far all a, e F and far all v
}
,v
2
e V.
The vector space V is called the domain of the transformation C while VV, the space into
which it maps, is called the co-domain.
Example 3.2.
1. Let F = R and take V = W = PC[f
0
, +00).
Define : PC[t
0
, +00) -> PC[t
0
, +00) by
2. Let F = R and take V = W = R
mx
". Fix M e R
mxm
.
Define : R
mx
" -> M
mxn
by
3. Let F = R and take V = P" = (p(x) = a
0
+ ct
}
x H h a
n
x" : a, E R} and
w = -p
n
-
1
.
Define C.: V > W by Lp p', where' denotes differentiation with respect to x.
17
Chapter 3
Linear Transformations
We begin with the basic definition of a linear transformation (or linear map, linear function,
or linear operator) between two vector spaces.
Definition 3.1. Let (V, IF) and (W, IF) be vector spaces. Then I:- : V -+ W is a linear
transformation if and only if
I:-(avi + {3V2) = al:-vi + {3I:-V2 for all a, {3 ElF and for all VI, V2 E V.
The vector space V is called the domain of the transformation I:- while W, the space into
which it maps, is called the co-domain.
Example 3.2.
1. Let IF = JR and take V = W = PC[to, +00).
Define I:- : PC[to, +00) -+ PC[to, +00) by
vet) f--+ wet) = (I:-v)(t) = 11 e-(t-r)v(r) dr.
to
2. Let IF = JR and take V = W = JRmxn. Fix ME JRmxm.
Define I:- : JRmxn -+ JRmxn by
X f--+ Y = I:-X = MX.
3. Let IF = JR and take V = pn = {p(x) = ao + alx + ... + anx
n
: ai E JR} and
W = pn-l.
Define I:- : V -+ W by I:- p = p', where I denotes differentiation with respect to x.
17
18 Chapters. Li near Transformations
Linear transformations between vector spaces with specific bases can be represented con-
veniently in matrix form. Specifically, suppose : (V, F) > (W, F) is linear and further
suppose that {u,, i e n} and {Wj, j e m] are bases for V and W, respectively. Then the
ith column of A = Mat (the matrix representation of with respect to the given bases
for V and W) is the representation of i> , with respect to {w
}
, j e raj. In other words,
represents since
where W = [w\,..., w
m
] and
is the z'th column of A. Note that A = Mat depends on the particular bases for V and W.
This could be reflected by subscripts, say, in the notation, but this is usually not done.
The action of on an arbitrary vector v e V is uniquely determined (by linearity)
by its action on a basis. Thus, if v = E1v1 + + E
n
v
n
= Vx (where u, and hence jc, is
arbitrary), then
Thinking of A both as a matrix and as a linear transformation from Rn to R
m
usually causes no
confusion. Change of basis then corresponds naturally to appropriate matrix multiplication.
Thus, V = WA since x was arbitrary.
When V = R", W = R
m
and [vi , i e n}, [wj , j e m} are the usual (natural) bases
the equation V = WA becomes simply = A. We thus commonly identify A as a linea
transformation with its matrix representation, i.e.,
18 Chapter 3. Linear Transformations
Linear transformations between vector spaces with specific bases can be represented con-
veniently in matrix form. Specifically, suppose L : (V, IF) (W, IF) is linear and further
suppose that {Vi, i E and {w j, j E !!!.} are bases for V and W, respectively. Then the
ith column of A = Mat L (the matrix representation of L with respect to the given bases
for V and W) is the representation of LVi with respect to {w j, j E m}. In other words,
represents L since
A=
al
n
]
: E JR.mxn
a
mn
LVi = aliwl + ... + amiWm
=Wai,
where W = [WI, ... , w
m
] and
is the ith column of A. Note that A = Mat L depends on the particular bases for V and W.
This could be reflected by subscripts, say, in the notation, but this is usually not done.
The action of L on an arbitrary vector V E V is uniquely determined (by linearity)
by its action on a basis. Thus, if V = VI + ... + Vn = V x (where v, and hence x, is
arbitrary), then
LVx = Lv = + ... +

= WAx.
Thus, LV = W A since x was arbitrary.
When V = JR.n, W = lR.
m
and {Vi, i E {W j' j E !!!.} are the usual (natural) bases,
the equation LV = W A becomes simply L = A. We thus commonly identify A as a linear
transformation with its matrix representation, i.e.,
Thinking of A both as a matrix and as a linear transformation from JR." to lR.
m
usually causes no
confusion. Change of basis then corresponds naturally to appropriate matrix multiplication.
3.3. Composition of Transformations 19
3.3 Composition of Transformations
Consider three vector spaces U, V, and W and transformations B from U to V and A from
V to W. Then we can define a new transformation C as follows:
formula
Two Special Cases:
Inner Product: Let x, y e Rn. Then their inner product is the scalar
Outer Product: Let x e R
m
, y e Rn. Then their outer product is the m x n
matrix
Note that any rank-one matrix A e R
mxn
can be written in the form A = xy
T
above (or xy
H
if A e C
mxn
). A rank-one symmetric matrix can be written in
the form XX
T
(or XX
H
).
The above diagram illustrates the composition of transformations C = AB. Note that in
most texts, the arrows above are reversed as follows:
However, it might be useful to prefer the former since the transformations A and B appear
in the same order in both the diagram and the equation. If dimZ// = p, dimV = n,
and dim W = m, and if we associate matrices with the transformations in the usual way,
then composition of transformations corresponds to standard matrix multiplication. That is,
we have C A B . The above is sometimes expressed componentwise by the
3.3. Composition ofTransformations 19
3.3 Composition of Transformations
Consider three vector spaces U, V, and Wand transformations B from U to V and A from
V to W. Then we can define a new transformation C as follows:
C
The above diagram illustrates the composition of transformations C = AB. Note that in
most texts, the arrows above are reversed as follows:
C
However, it might be useful to prefer the former since the transformations A and B appear
in the same order in both the diagram and the equation. If dimU = p, dim V = n,
and dim W = m, and if we associate matrices with the transformations in the usual way,
then composition of transformations corresponds to standard matrix mUltiplication. That is,
we have CAB . The above is sometimes expressed componentwise by the
mxp
formula
Two Special Cases:
nxp
n
cij = L aikbkj.
k=1
Inner Product: Let x, y E ~ n . Then their inner product is the scalar
n
xTy = Lx;y;.
;=1
Outer Product: Let x E ~ m , y E ~ n . Then their outer product is the m x n
matrix
Note that any rank-one matrix A E ~ m x n can be written in the form A = xyT
above (or xyH if A E c
mxn
). A rank-one symmetric matrix can be written in
the form xx
T
(or xx
H
).
20 Chapter 3. Li near Transformations
Let A : V > W be a linear transformation.
Definition 3.3. The range of A, denotedlZ( A), is the set {w e W : w = Av for some v e V}.
Equivalently, R(A) {Av : v e V}. The range of A is also known as the image of A and
denoted Im(A).
The nullspace of A, denoted N(A), is the set {v e V : Av = 0}. The nullspace of
A is also known as the kernel of A and denoted Ker (A).
Theorem 3.4. Let A : V > W be a linear transformation. Then
1. R( A) C W.
2. N(A) c V.
Note that N(A) and R(A) are, in general, subspaces of different spaces.
Theorem 3.5. Let A e R
mxn
. If A is written in terms of its columns as A = [a\,... ,a
n
],
then
Proof: The proof of this theorem is easy, essentially following immediately from the defi-
nition. D
Remark 3.6. Note that in Theorem 3.5 and throughout the text, the same symbol (A) is
used to denote both a linear transformation and its matrix representation with respect to the
usual (natural) bases. See also the last paragraph of Section 3.2.
Definition 3.7. Let {v1 , . . . , vk] be a set of nonzero vectors u, e Rn. The set is said to
be orthogonal if' vjvj = 0 for i ^ j and orthonormal if vf vj = 8
ij
, where 8
t
j is the
Kronecker delta defined by
Example 3.8.
is an orthogonal set.
is an orthonormal set.
3. If { t > i , . . . , Vk} with u, M." is an orthogonal set, then I /==, - -., /=== | is an
I ^/v, vi ^/v'
k
v
k
]
orthonormal set.
then
20 Chapter 3. LinearTransformations
Let A : V --+ W be a linear transformation.
Definition3.3. The range of A, denotedR(A), is the set {w E W : w = Av for some v E V}.
Equivalently, R(A) = {Av : v E V}. The range of A is also known as the image of A and
denoted Im(A).
The nullspace of A, denoted N(A), is the set {v E V : Av = O}. The nullspace of
A is also known as the kernel of A and denoted Ker (A).
Theorem 3.4. Let A : V --+ W be a linear transformation. Then
1. R(A) S; W.
2. N(A) S; V.
Note that N(A) and R(A) are, in general, subspaces of different spaces.
Theorem 3.5. Let A E If A is written in terms of its columns as A = [ai, ... , an],
then
R(A) = Sp{al, ... , an} .
Proof: The proof of this theorem is easy, essentially following immediately from the defi-
nition. 0
Remark 3.6. Note that in Theorem 3.5 and throughout the text, the same symbol (A) is
used to denote both a linear transformation and its matrix representation with respect to the
usual (natural) bases. See also the last paragraph of Section 3.2.
Definition 3.7. Let {VI, ... , vd be a set of nonzero vectors Vi E The set is said to
be orthogonal if vr v j = 0 for i f= j and orthonormal if vr v j = 8ij' where 8ij is the
Kronecker delta defined by
8 {I ifi=j,
ij = 0 if i f= j.
Example 3.8.
1. {[ J. [ -: J} is an orthogonal set.
2. {[ ] ,[ J} is an orthonormal set.
3 If { }
. h 1Tlln h I th { .
. VI, . ,Vk Wit Vi E.IN,. IS an ort ogona set, en ... , IS an
VI
orthonormal set.
3.4. Structure of Linear Transformations 21
Definition 3.9. Let S c Rn. Then the orthogonal complement of S is defined as the set
S
1
- = {v e Rn : V
T
S = 0 for all s e S}.
Example 3.10. Let
Then it can be shown that
Working from the definition, the computation involved is simply to find all nontrivial (i.e.,
nonzero) solutions of the system of equations
Note that there is nothing special about the two vectors in the basis defining S being or-
thogonal. Any set of vectors will do, including dependent spanning vectors (which would,
of course, then give rise to redundant equations).
Theorem 311 Let R S C R
n
The
Proof: We prove and discuss only item 2 here. The proofs of the other results are left as
exercises. Let { v1 , ..., v
k
} be an orthonormal basis for S and let x e Rn be an arbitrary
vector. Set
3.4. Structure of Li near Transformations 21
Definition 3.9. Let S <; ]Rn. Then the orthogonal complement of S is defined as the set
vTs=OforallsES}.
Example 3.10. Let
Then it can be shown that
Working from the definition, the computation involved is simply to find all nontrivial (i.e.,
nonzero) solutions of the system of equations
3xI + 5X2 + 7X3 = 0,
-4xI + X2 + X3 = 0.
Note that there is nothing special about the two vectors in the basis defining S being or-
thogonal. Any set of vectors will do, including dependent spanning vectors (which would,
of course, then give rise to redundant equations).
Theorem 3.11. Let n, S <; ]Rn. Then
2. S \B = ]Rn.
3. = S.
4. n <; S if and only if <;
5. (n + = nl. n
6. (n n = +
Proof: We prove and discuss only item 2 here. The proofs of the other results are left as
exercises. Let {VI, ... , Vk} be an orthonormal basis for S and let x E ]Rn be an arbitrary
vector. Set
k
XI = L (xT Vi)Vi,
;=1
X2 = X -XI.
we see that x2 is orthogonal to v1, ..., Vk and hence to any linear combination of these
vectors. In other words, X2 is orthogonal to any vector in S. We have thus shown that
S + S
1
= Rn. We also have that S U S
1
=0 since the only vector s e S orthogonal to
everything in S (i.e., including itself) is 0.
It is also easy to see directly that, when we have such direct sum decompositions, we
can write vectors in a unique way with respect to the corresponding subspaces. Suppose,
for example, that x = x1 + x2. = x'1+ x'
2
, where x\, x 1 E S and x2, x'
2
e S
1
. Then
(x'1 x1)
T
( x'
2
x2) = 0 by definition of ST . But then (x'1 x1)
T
( x' 1 x1) = 0 since
x
2
X2 = (x'1 x1) (which follows by rearranging the equation x1+x2 = x'1 + x'
2
) . Thus,
x1 x'1 and x2 = x
2
. D
Theorem 3.12. Let A : Rn > R
m
. Then
1. N(A)
1
" = 7(A
r
). (Note: This holds only for finite-dimensional vector spaces.)
2. 'R,(A)
1
~ J\f(A
T
). (Note: This also holds for infinite-dimensional vector spaces.)
Proof: To prove the first part, take an arbitrary x e A/ "(A). Then Ax = 0 and this is
equivalent to y
T
Ax = 0 for all v. But y
T
Ax = ( A
T
y ) x. Thus, Ax = 0 if and only if x
is orthogonal to all vectors of the form A
T
v, i.e., x e R(A
r
) . Since x was arbitrary, we
have established that N(A)
1
= U(A
T
}.
The proof of the second part is similar and is left as an exercise. D
Definition 3.13. Let A : R
n
-> R
m
. Then {v e R" : Av = 0} is sometimes called the
right nullspace of A. Similarly, (w e R
m
: W
T
A = 0} is called the left nullspace of A.
Clearly, the right nullspace is A/"(A) while the left nullspace is J\f(A
T
).
Theorem 3.12 and part 2 of Theorem 3.11 can be combined to give two very fun-
damental and useful decompositions of vectors in the domain and co-domain of a linear
transformation A. See also Theorem 2.26.
Theorem 3.14 (Decomposition Theorem). Let A : R" -> R
m
. Then
7. every vector v in the domain space R" can be written in a unique way as v = x + y,
where x M(A) and y J\f(A)
= ft(A
r
) (i.e., R" = M(A) 0 ft(A
r
)).
2. every vector w in the co-domain space R
m
can be written in a unique way asw = x+y,
where x e U(A) and y e ft(A)
1
- = Af(A
T
) (i.e., R
m
= 7l(A) 0 M(A
T
)).
This key theorem becomes very easy to remember by carefully studying and under-
standing Figure 3.1 in the next section.
3.5 Four Fundamental Subspaces
Consider a general matrix A E^
x
". When thought of as a linear transformation from E"
to R
m
, many properties of A can be developed in terms of the four fundamental subspaces
22 Chapters. L i near Transformations
Then x\ e < S and, since
Then XI E S and, since
T T T
x
2
V j = X V j - X I V j
=XTVj-XTVj=O,
we see that X2 is orthogonal to VI, .. , Vk and hence to any linear combination of these
vectors. In other words, X2 is orthogonal to any vector in S. We have thus shown that
S + S.l = IRn. We also have that S n S.l = 0 since the only vector s E S orthogonal to
everything in S (i.e., including itself) is O.
It is also easy to see directly that, when we have such direct sum decompositions, we
can write vectors in a unique way with respect to the corresponding subspaces. Suppose,
for example, that x = XI + X2 = x; + ~ , where XI, x; E Sand X2, x ~ E S.l. Then
(x; - XI/ x ~ - X2) = 0 by definition of S.l. But then (x; - XI)T (x; - xd = 0 since
x ~ -X2 = -(x; -XI) (which follows by rearranging the equation XI +X2 = x; + x ~ ) . Thus,
XI = x; andx2 = x ~ . 0
Theorem 3.12. Let A : IR
n
-+ IRm. Then
1. N(A).l = R(A
T
). (Note: This holds only for finite-dimensional vector spaces.)
2. R(A).l = N(A
T
). (Note: This also holds for infinite-dimensional vector spaces.)
Proof: To prove the first part, take an arbitrary x E N(A). Then Ax = 0 and this is
equivalent to yT Ax = 0 for all y. But yT Ax = (AT y{ x. Thus, Ax = 0 if and only if x
is orthogonal to all vectors of the form AT y, i.e., x E R(AT).l. Since x was arbitrary, we
have established thatN(A).l = R(A
T
).
The proof of the second part is similar and is left as an exercise. 0
Definition 3.13. Let A : IR
n
-+ IRm. Then {v E IR
n
: A v = O} is sometimes called the
right nullspace of A. Similarly, {w E IR
m
: w
T
A = O} is called the left nullspace of A.
Clearly, the right nullspace is N(A) while the left nullspace is N(A
T
).
Theorem 3.12 and part 2 of Theorem 3.11 can be combined to give two very fun-
damental and useful decompositions of vectors in the domain and co-domain of a linear
transformation A. See also Theorem 2.26.
Theorem 3.14 (Decomposition Theorem). Let A : IR
n
-+ IRm. Then
1. every vector v in the domain space IR
n
can be written in a unique way as v = x + y,
where x E N(A) and y E N(A).l = R(AT) (i.e., IR
n
= N(A) EB R(A
T
.
2. every vector w in the co-domain space IR
m
can be written ina unique way as w = x+y,
where x E R(A) and y E R(A).l = N(A
T
) (i.e., IR
m
= R(A) EBN(A
T
.
This key theorem becomes very easy to remember by carefully studying and under-
standing Figure 3.1 in the next section.
3.5 Four Fundamental Subspaces
Consider a general matrix A E lR;,xn. When thought of as a linear transformation from IR
n
to IRm, many properties of A can be developed in terms of the four fundamental subspaces
3.5. Four Fundamental Subspaces 23
Figure 3.1. Four fundamental subspaces.
7(A), 'R.(A)^, Af ( A) , and N(A)T. Figure 3.1 makes many key properties seem almost
obvious and we return to this figure frequently both in the context of linear transformations
and in illustrating concepts such as controllability and observability.
Definition 3.15. Let V and W be vector spaces and let A : V
motion.
1. A is onto (also called epic or surjective) ifR,(A) = W.
W be a linear transfor-
2. A is one-to-one or 1-1 (also called monic or infective) ifJ\f(A) = 0. Two equivalent
characterizations of A being 1-1 that are often easier to verify in practice are the
following:
Definition 3.16. Let A : E" -> R
m
. Then rank(A) = dimftCA). This is sometimes called
the column rank of A (maximum number of independent columns). The row rank of A is
A
r
N(A)1-
r
EB {OJ
X {O}Gl
n-r m -r
Figure 3.1. Four fundamental subspaces.
R(A), R(A)1-, N(A), and N(A)1-. Figure 3.1 makes many key properties seem almost
obvious and we return to this figure frequently both in the context of linear transformations
and in illustrating concepts such as controllability and observability.
Definition 3.15. Let V and W be vector spaces and let A : V -+ W be a linear transfor-
mation.
1. A is onto (also called epic or surjective) ifR(A) = W.
2. A is one-to-one or 1-1 (also called monic or injective) if N(A) = O. Two equivalent
characterizations of A being 1-1 that are often easier to verify in practice are the
following:
(a) AVI = AV2 ===} VI = V2 .
(b) VI t= V2 ===} AVI t= AV2 .
Definition 3.16. Let A : IR
n
-+ IRm. Then rank(A) = dim R(A). This is sometimes called
the column rank of A (maximum number of independent columns). The row rank of A is
24 Chapter3. Linear Transformations
dim 7(A
r
) (maximum number of independent rows). The dual notion to rank is the nullity
of A, sometimes denoted nullity(A) or corank(A), and is defined as dim A/"(A).
Theorem 3.17. Let A : R
n
-> R
m
. Then dim K(A) = dimA/ '(A)
. (Note: Since
A/Â)
1
" = 7l(A
T
), this theorem is sometimes colloquially stated "row rank of A = column
rank of A.")
Proof: Define a linear transformation T : J\f(A)~
L
> 7(A) by
Clearly T is 1-1 (since A/"(T) = 0). To see that T is also onto, take any w e 7(A). Then
by definition there is a vector x e R" such that Ax w. Write x = x\ + X2, where
x\ e A/Â)
1
- and jc
2
e A/"(A). Then Ajti = u; = r*i since *i e A/Â)-
1
. The last equality
shows that T is onto. We thus have that dim7?.(A) = dimA/Â^ since it is easily shown
that if { ui , . . . , iv} is abasis forA/'CA)
1
, then {Tv\, . . . , Tv
r
] is abasis for 7?.(A). Finally, if
we apply this and several previous results, the following string of equalities follows easily:
"column rank of A" = rank(A) = dim7e(A) = dim A/Â)
1
= dim7l(A
T
) = rank(A
r
) =
"row rank of A." D
The following corollary is immediate. Like the theorem, it is a statement about equality
of dimensions; the subspaces themselves are not necessarily in the same vector space.
Corollary 3.18. Let A : R" -> R
m
. Then dimA/"(A) + dimft(A) = n, where n is the
dimension of the domain of A.
Proof: From Theorems 3.11 and 3.17 we see immediately that
For completeness, we include here a few miscellaneous results about ranks of sums
and products of matrices.
Theorem 3.19. Let A, B e R"
xn
. Then
Part 4 of Theorem 3.19 suggests looking at the general problem of the four fundamental
subspaces of matrix products. The basic results are contained in the following easily proved
theorem.
24 Chapter 3. LinearTransformations
dim R(AT) (maximum number of independent rows). The dual notion to rank is the nullity
of A, sometimes denoted nullity(A) or corank(A), and is defined as dimN(A).
Theorem 3.17. Let A : ]Rn ~ ]Rm. Then dim R(A) = dimNCA)-L. (Note: Since
N(A)-L = R(A
T
), this theorem is sometimes colloquially stated "row rank of A = column
rank of A.")
Proof: Define a linear transformation T : N(A)-L ~ R(A) by
Tv = Av for all v E N(A)-L.
Clearly T is 1-1 (since N(T) = 0). To see that T is also onto, take any W E R(A). Then
by definition there is a vector x E ]Rn such that Ax = w. Write x = Xl + X2, where
Xl E N(A)-L andx2 E N(A). Then AXI = W = TXI since Xl E N(A)-L. The last equality
shows that T is onto. We thus have that dim R(A) = dimN(A)-L since it is easily shown
that if {VI, ... , v
r
} is a basis for N(A)-L, then {TVI, ... , Tv
r
} is a basis for R(A). Finally, if
we apply this and several previous results, the following string of equalities follows easily:
"column rank of A" = rank(A) = dim R(A) = dimN(A)-L = dim R(AT) = rank(AT) =
"row rank of A." 0
The following corollary is immediate. Like the theorem, it is a statement about equality
of dimensions; the subspaces themselves are not necessarily in the same vector space.
Corollary 3.18. Let A : ]Rn ~ ]Rm. Then dimN(A) + dim R(A) = n, where n is the
dimension of the domain of A.
Proof: From Theorems 3.11 and 3.17 we see immediately that
n = dimN(A) + dimN(A)-L
= dimN(A) + dim R(A) . 0
For completeness, we include here a few miscellaneous results about ranks of sums
and products of matrices.
Theorem 3.19. Let A, B E ]Rnxn. Then
1. O:s rank(A + B) :s rank(A) + rank(B).
2. rank(A) + rank(B) - n :s rank(AB) :s min{rank(A), rank(B)}.
3. nullity(B) :s nullity(AB) :s nullity(A) + nullity(B).
4. if B is nonsingular, rank(AB) = rank(BA) = rank(A) and N(BA) = N(A).
Part 4 of Theorem 3.19 suggests looking atthe general problem of the four fundamental
subspaces of matrix products. The basic results are contained in the following easily proved
theorem.
3.5. Four F u n d a me n t a l Subspaces 25
mxn
, B e R
nxp
. Then
The next theorem is closely related to Theorem 3.20 and is also easily proved. It
is extremely useful in text that follows, especially when dealing with pseudoinverses and
linear least squares problems.
mxn
. Then
We now characterize 1-1 and onto transformations and provide characterizations in
terms of rank and invertibility.
Theorem 3.22. Let A : R
n
- R
m
. Then
1. A is onto if and only //"rank(A) m (A has linearly independent rows or is said to
have full row rank; equivalently, AA
T
is nonsingular).
2. A is 1-1 if and only z/r a nk(A) = n (A has linearly independent columns or is said
to have full column rank; equivalently, A
T
A is nonsingular).
Proof: Proof of part 1: If A is onto, dim7?,(A) m rank (A). Conversely, let y e R
m
be arbitrary. Let jc = A
T
(AA
T
)~
]
y e R
n
. Then y = Ax, i.e., y e 7?.(A), so A is onto.
Proof of part 2: If A is 1-1, then A/"(A) = 0, which implies that dim A/Â)-
1
n
dim 7(A
r
), and hence dim 7(A) = n by Theorem 3.17. Conversely, suppose Ax\ = Ax^.
Then A
r
A;t i = A
T
Ax2, which implies x\ = x^. since A
r
A is invertible. Thus, A is
1-1. D
Definition 3.23. A : V W is invertible (or bijective) if and only if it is 1-1 and onto.
Note that if A is invertible, then dim V dim W. Also, A : W
1
- E" is invertible or
nonsingular if and only z/r ank(A) = n.
Note that in the special case when A R"
x
", the transformations A, A
r
, and A"
1
are all 1-1 and onto between the two spaces M(A)
and 7(A). The transformations A

T
and A~
!
have the same domain and range but are in general different maps unless A is
orthogonal. Similar remarks apply to A and A~
T
.
Theorem 3.20. Let A E IRmxn, B E IRnxp. Then
1. RCAB) S; RCA).
2. N(AB) ;2 N(B).
3. RAB)T) S; R(B
T
).
4. NAB)T) ;2 N(A
T
).
The next theorem is closely related to Theorem 3.20 and is also easily proved. It
is extremely useful in text that follows, especially when dealing with pseudoinverses and
linear least squares problems.
Theorem 3.21. Let A E IRmxn. Then
1. R(A) = R(AA
T
).
2. R(AT) = R(A
T
A).
3. N(A) = N(A
T
A).
4. N(A
T
) = N(AA
T
).
We now characterize I-I and onto transformations and provide characterizations in
terms of rank and invertibility.
Theorem 3.22. Let A : IR
n
-+ IRm. Then
1. A is onto if and only if rank (A) = m (A has linearly independent rows or is said to
have full row rank; equivalently, AA T is nonsingular).
2. A is 1-1 if and only ifrank(A) = n (A has linearly independent columns or is said
to have full column rank; equivalently, AT A is nonsingular).
Proof' Proof of part 1: If A is onto, dim R(A) = m = rank(A). Conversely, let y E IRm
be arbitrary. Let x = AT (AAT)-I Y E IRn. Then y = Ax, i.e., y E R(A), so A is onto.
Proof of part 2: If A is 1-1, then N(A) = 0, which implies that dimN(A)1- = n =
dim R(A
T
), and hence dim R(A) = n by Theorem 3.17. Conversely, suppose AXI = AX2.
Then AT AXI = AT AX2, which implies XI = X2 since AT A is invertible. Thus, A is
1-1. D
Definition 3.23. A : V -+ W is invertible (or bijective) if and only if it is 1-1 and onto.
Note that if A is invertible, then dim V = dim W. Also, A : IRn -+ IR
n
is invertible or
nonsingular ifand only ifrank(A) = n.
Note that in the special case when A E I R ~ x n , the transformations A, AT, and A-I
are all 1-1 and onto between the two spaces N(A)1- and R(A). The transformations AT
and A -I have the same domain and range but are in general different maps unless A is
orthogonal. Similar remarks apply to A and A -T.
26 Chapters. Li near Transformations
If a linear transformation is not invertible, it may still be right or left invertible. Defi-
nitions of these concepts are followed by a theorem characterizing left and right invertible
transformations.
Definition 3.24. Let A : V -> W. Then
1. A is said to be right invertible if there exists a right inverse transformation A~
R
:
W > V such that AA~
R
= I
w
, where I
w
denotes the identity transformation on W.
2. A is said to be left invertible if there exists a left inverse transformation A~
L
: W >
V such that A~
L
A = I
v
, where I
v
denotes the identity transformation on V.
Theorem 3.25. Let A : V -> W. Then
1. A is right invertible if and only if it is onto.
2. A is left invertible if and only if it is 1-1.
Moreover, A is invertible if and only if it is both right and left invertible, i.e., both 1-1 and
onto, in which case A~
l
= A~
R
= A~
L
.
Note: From Theorem 3.22 we see that if A : E" -> E
m
is onto, then a right inverse
is given by A~
R
= A
T
(AA
T
) . Similarly, if A is 1-1, then a left inverse is given by
A~
L
= (A
T
A)~
1
A
T
.
Theorem 3.26. Let A : V - V.
1. If there exists a unique right inverse A~
R
such that AA~
R
= I, then A is invertible.
2. If there exists a unique left inverse A~
L
such that A~
L
A = I, then A is invertible.
Proof: We prove the first part and leave the proof of the second to the reader. Notice the
following:
Thus, (A
R
+ A
R
A /) must be a right inverse and, therefore, by uniqueness it must be
the case that A~
R
+ A~
R
A I = A~
R
. But this implies that A~
R
A = /, i.e., that A~
R
is
a left inverse. It then follows from Theorem 3.25 that A is invertible. D
Example 3.27.
1. Let A = [1 2] : E
2
- E
1
. Then A is onto. (Proof: Take any a E
1
; then one
can always find v e E
2
such that [1 2][^] = a). Obviously A has full row rank
(=1) and A~
R
= [ _j j is a right inverse. Also, it is clear that there are infinitely many
right inverses for A. In Chapter 6 we characterize all right inverses of a matrix by
characterizing all solutions of the linear matrix equation AR = I.
26 Chapter 3. linear Transformations
If a linear transformation is not invertible, it may still be right or left invertible. Defi-
nitions of these concepts are followed by a theorem characterizing left and right invertible
transformations.
Definition 3.24. Let A : V -+ W. Then
1. A is said to be right invertible if there exists a right inverse transformation A-
R
:
W -+ V such that AA -R = I
w
, where Iw denotes the identity transfonnation on W.
2. A is said to be left invertible if there exists a left inverse transformation A -L : W -+
V such that A -L A = Iv, where Iv denotes the identity transfonnation on V.
Theorem 3.25. Let A : V -+ W. Then
1. A is right invertible if and only if it is onto.
2. A is left invertible if and only ifit is 1-1.
Moreover, A is invertible if and only if it is both right and left invertible, i.e., both 1-1 and
onto, in which case A -I = A -R = A -L.
Note: From Theorem 3.22 we see that if A : ]Rn -+ ]Rm is onto, then a right inverse
is given by A -R = AT (AAT) -I. Similarly, if A is 1-1, then a left inverse is given by
A-
L
= (AT A)-I AT.
Theorem 3.26. Let A : V -+ V.
1. If there exists a unique right inverse A - R such that A A - R = I, then A is invertible.
2. If there exists a unique left inverse A -L such that A -L A = I, then A is invertible.
Proof: We prove the first part and leave the proof of the second to the reader. Notice the
following:
A(A-
R
+ A-RA -I) = AA-
R
+ AA-RA - A
= I + I A - A since AA -R = I
= I.
Thus, (A -R + A -R A - I) must be a right inverse and, therefore, by uniqueness it must be
the case that A -R + A -R A - I = A -R. But this implies that A -R A = I, i.e., that A -R is
a left inverse. It then follows from Theorem 3.25 that A is invertible. 0
Example 3.27.
1. Let A = [1 2]:]R2 -+ ]R I. Then A is onto. (Proo!' Take any a E ]R I; then one
can always find v E ]R2 such that [1 2][ ~ ] = a). Obviously A has full row rank
(= 1) and A - R = [ _ ] is a right inverse. Also, it is clear that there are infinitely many
right inverses for A. In Chapter 6 we characterize all right inverses of a matrix by
characterizing all solutions of the linear matrix equation A R = I.
Exercises 27
2. Let A = [J] : E
1
-> E
2
. ThenAis 1-1. (Proof: The only solution to 0 = Av = [
I
2
]v
is v = 0, whence A/"(A) = 0 so A is 1-1). It is now obvious that A has full column
rank (=1) and A~
L
= [3 1] is a left inverse. Again, it is clear that there are
infinitely many left inverses for A. In Chapter 6 we characterize all left inverses of a
matrix by characterizing all solutions of the linear matrix equation LA = I.
3. The matrix
when considered as a linear transformation on IE
below bases for its four fundamental subspaces.
\ is neither 1-1 nor onto. We give
EXERCISES
3 4
1. Let A = [
8 5
J and consider A as a linear transformation mapping E
3
to E
2
.
Find the matrix representation of A with respect to the bases
2. Consider the vector space R
nx
" over E, let S denote the subspace of symmetric
matrices, and let 7 denote the subspace of skew-symmetric matrices. For matrices
X, Y e E
nx
" define their inner product by (X, Y) = Tr( X
r
F) . Show that, with
respect to this inner product, 'R, S^.
3. Consider the differentiation operator C defined in Example 3.2.3. Is 1-1? Is
onto?
4. Prove Theorem 3.4.
of R
3
and
of E
2
.
Exercises 27
2. LetA = [i]:]Rl ~ ]R2. Then A is 1-1. (Proof The only solution toO = Av = [i]v
is v = 0, whence N(A) = 0 so A is 1-1). It is now obvious that A has full column
rank (=1) and A -L = [3 - 1] is a left inverse. Again, it is clear that there are
infinitely many left inverses for A. In Chapter 6 we characterize all left inverses of a
matrix by characterizing all solutions of the linear matrix equation LA = I.
3. The matrix
[
1 1
A = 2 1
3 1
when considered as a linear transformation on ]R3, is neither 1-1 nor onto. We give
below bases for its four fundamental subspaces.
EXERCISES
1. Let A = [ ~ ; i) and consider A as a linear transformation mapping ]R3 to ]R2.
Find the matrix representation of A with respect to the bases
{[lHHU]}
{ [ i l [ ~ J }
2. Consider the vector space ]Rnxn over ]R, let S denote the subspace of symmetric
matrices, and let R denote the subspace of skew-symmetric matrices. For matrices
X, Y E ]Rnxn define their inner product by (X, y) = Tr(X
T
Y). Show that, with
respect to this inner product, R = S J. .
3. Consider the differentiation operator , defined in Example 3.2.3. Is , I-I? Is ,
onto?
28 Chapters. Linear Transformations
5. Prove Theorem 3.11.4.
7. Determine bases for the four fundamental subspaces of the matrix
8. Suppose A e R
mxn
has a left inverse. Show that A
T
has a right inverse.
9. Let A = [ J o]. Determine A/"(A) and 7(A). Are they equal? Is this true in general?
If this is true in general, prove it; if not, provide a counterexample.
10. Suppose A Mg
9x48
. How many linearly independent solutions can be found to the
homogeneous linear system Ax = 0?
11. Modify Figure 3.1 to illustrate the four fundamental subspaces associated with A
T
e
R
nxm
thought of as a transformation from R
m
to R".
5. Prove Theorem 3.Il.4.
7. Detennine bases for the four fundamental subspaces of the matrix

2 5 5 3
8. Suppose A E IR
m
xn has a left inverse. Show that A T has a right inverse.
9. Let A = n DetennineN(A) and R(A). Are they equal? Is this true in general?
If this is true in general, prove it; if not, provide a counterexample.
10. Suppose A E How many linearly independent solutions can be found to the
homogeneous linear system Ax = O?
11. Modify Figure 3.1 to illustrate the four fundamental subspaces associated with ATE
IR
nxm
thought of as a transformation from IR
m
to IRn.
Chapter 4
Introduction to the
Moore-Pen rose
Pseudoinverse
In this chapter we give a brief introduction to the Moore-Penrose pseudoinverse, a gener-
alization of the inverse of a matrix. The Moore-Penrose pseudoinverse is defined for any
matrix and, as is shown in the following text, brings great notational and conceptual clarity
to the study of solutions to arbitrary systems of linear equations and linear least squares
problems.
4.1 Definitions and Characterizations
Consider a linear transformation A : X > y, where X and y are arbitrary finite-
dimensional vector spaces. Define a transformation T : Af(A)
1
- > Tl(A) by
Then, as noted in the proof of Theorem 3.17, T is bijective (1-1 and onto), and hence we
can define a unique inverse transformation T~
l
: 7(A) > J\f(A}~
L
. This transformation
can be used to give our first definition of A
+
, the Moore-Penrose pseudoinverse of A.
Unfortunately, the definition neither provides nor suggests a good computational strategy
for determining A
+
.
Definition 4.1. With A and T as defined above, define a transformation A
+
: y X by
where y = y\ + j2 with y\ e 7(A) and yi e Tl(A}
L
. Then A
+
is the Moore-Penrose
pseudoinverse of A.
Although X and y were arbitrary vector spaces above, let us henceforth consider the
case X = W
1
and y = R
m
. We have thus defined A+ for all A e IR
X
" . A purely algebraic
characterization of A
+
is given in the next theorem, which was proved by Penrose in 1955;
see [22].
29
Chapter 4
Introduction to the
Moore-Penrose
Pseudoinverse
In this chapter we give a brief introduction to the Moore-Penrose pseudoinverse, a gener-
alization of the inverse of a matrix. The Moore-Penrose pseudoinverse is defined for any
matrix and, as is shown in the following text, brings great notational and conceptual clarity
to the study of solutions to arbitrary systems of linear equations and linear least squares
problems.
4.1 Definitions and Characterizations
Consider a linear transformation A : X ---+ y, where X and Y are arbitrary finite-
dimensional vector spaces. Define a transformation T : N(A).l ---+ R(A) by
Tx = Ax for all x E NCA).l.
Then, as noted in the proof of Theorem 3.17, T is bijective Cl-l and onto), and hence we
can define a unique inverse transformation T-
1
: RCA) ---+ NCA).l. This transformation
can be used to give our first definition of A +, the Moore-Penrose pseudoinverse of A.
Unfortunately, the definition neither provides nor suggests a good computational strategy
for determining A + .
Definition 4.1. With A and T as defined above, define a transformation A + : Y ---+ X by
where Y = YI + Yz with Yl E RCA) and Yz E RCA).l. Then A+ is the Moore-Penrose
pseudoinverse of A.
Although X and Y were arbitrary vector spaces above, let us henceforth consider the
case X = n and Y = lP1.
m
. We have thus defined A + for all A E lP1.;" xn. A purely algebraic
characterization of A + is given in the next theorem, which was proved by Penrose in 1955;
see [22].
29
30 Chapter 4. Introduction to the Moore-Penrose Pseudoinverse
Theorem 4.2. Let A e R?
xn
. Then G = A
+
i f and only i f
(PI) AGA = A.
(P2) GAG = G.
(P3) (AGf = AG.
(P4) (GA)
T
= GA.
Furthermore, A
+
always exi sts and i s uni que.
Note that the inverse of a nonsingular matrix satisfies all four Penrose properties. Also,
a right or left inverse satisfies no fewer than three of the four properties. Unfortunately, as
with Definition 4.1, neither the statement of Theorem 4.2 nor its proof suggests a computa-
tional algorithm. However, the Penrose properties do offer the great virtue of providing a
checkable criterion in the following sense. Given a matrix G that is a candidate for being
the pseudoinverse of A, one need simply verify the four Penrose conditions (P1)-(P4). If G
satisfies all four, then by uniqueness, it must be A
+
. Such a verification is often relatively
straightforward.
Example 4.3. Consider A = [' ]. Verify directly that A
+
= [| f ] satisfies (P1)-(P4).
Note that other left inverses (for example, A~
L
= [3 1]) satisfy properties (PI), (P2),
and (P4) but not (P3).
Still another characterization of A
+
is given in the following theorem, whose proof
can be found in [1, p. 19]. While not generally suitable for computer implementation, this
characterization can be useful for hand calculation of small examples.
xn
. Then
4.2 Examples
Each of the following can be derived or verified by using the above definitions or charac-
terizations.
Example 4.5. A
+
= A
T
(AA
T
)~ if A is onto (independent rows) (A is right invertible).
Example 4.6. A
+
= (A
T
A)~ A
T
i f A is 1-1 (independent columns) (A is left invertible).
Example 4.7. For any scalar a,
Theorem 4.2. Let A E lR;" xn. Then G = A + if and only if
(Pl) AGA = A.
(P2) GAG = G.
(P3) (AG)T = AG.
(P4) (GA)T = GA.
Furthermore, A + always exists and is unique.
Note that the inverse of a nonsingular matrix satisfies all four Penrose properties. Also,
a right or left inverse satisfies no fewer than three of the four properties. Unfortunately, as
with Definition 4.1, neither the statement of Theorem 4.2 nor its proof suggests a computa-
tional algorithm. However, the Penrose properties do offer the great virtue of providing a
checkable criterion in the following sense. Given a matrix G that is a candidate for being
the pseudoinverse of A, one need simply verify the four Penrose conditions (P1)-(P4). If G
satisfies all four, then by uniqueness, it must be A +. Such a verification is often relatively
straightforward.
Example 4.3. Consider A = [a Verify directly that A+ = [! ~ ] satisfies (PI)-(P4).
Note that other left inverses (for example, A -L = [3 - 1]) satisfy properties (PI), (P2),
and (P4) but not (P3).
Still another characterization of A + is given in the following theorem, whose proof
can be found in [1, p. 19]. While not generally suitable for computer implementation, this
characterization can be useful for hand calculation of small examples.
Theorem 4.4. Let A E lR;" xn. Then
4.2 Examples
A + = lim (AT A + 8
2
1) -I AT
6--+0
= limAT(AAT +8
2
1)-1.
6--+0
(4.1)
(4.2)
Each of the following can be derived or verified by using the above definitions or charac-
terizations.
Example 4.5. X
t
= AT (AA T) -I if A is onto (independent rows) (A is right invertible).
Example 4.6. A+ = (AT A)-I AT if A is 1-1 (independent columns) (A is left invertible).
Example 4.7. For any scalar a,
if a t= 0,
if a =0.
4.3. Properties and Appl ications 31
Example 4.8. For any vector v e M",
Example 4.9.
Example 4.10.
4.3 Properties and Applications
This section presents some miscellaneous useful results on pseudoinverses. Many of these
are used in the text that follows.
mx
" and suppose U e R
mxm
, V e R
nx
" are orthogonal (M is
orthogonal if M
T
= M
-1
). Then
Proof: For the proof, simply verify that the expression above does indeed satisfy each c
the four Penrose conditions. D
Theorem 4.12. Let S e R
nxn
be symmetric with U
T
SU = D, where U is orthogonal an
D is diagonal. Then S
+
= UD
+
U
T
, where D
+
is again a diagonal matrix whose diagonc
elements are determined according to Example 4.7.
Theorem 4.13. For all A e R
mxn
,
Proof: Both results can be proved using the limit characterization of Theorem 4.4. The
proof of the first result is not particularly easy and does not even have the virtue of being
especially illuminating. The interested reader can consult the proof in [1, p. 27]. The
proof of the second result (which can also be proved easily by verifying the four Penrose
conditions) is as follows:
4.3. Properties and Applications
Example 4.8. For any vector v E jRn,
Example 4.9.
[ ~ ~ r = [
0
~ l
[ ~ r = [
I I
1
Example 4.10.
4 4
I I
4 4
4.3 Properties and Applications
if v i= 0,
if v = O.
31
This section presents some miscellaneous useful results on pseudoinverses. Many of these
are used in the text that follows.
Theorem 4.11. Let A E jRmxn and suppose U E jRmxm, V E jRnxn are orthogonal (M is
orthogonal if MT = M-
1
). Then
Proof: For the proof, simply verify that the expression above does indeed satisfy each of
the four Penrose conditions. 0
Theorem 4.12. Let S E jRnxn be symmetric with U
T
SU = D, where U is orthogonal and
D is diagonal. Then S+ = U D+U
T
, where D+ is again a diagonal matrix whose diagonal
elements are determined according to Example 4.7.
Theorem 4.13. For all A E jRmxn,
1. A+ = (AT A)+ AT = AT (AA
T
)+.
2. (A
T
)+ = (A+{.
Proof: Both results can be proved using the limit characterization of Theorem 4.4. The
proof of the first result is not particularly easy and does not even have the virtue of being
especially illuminating. The interested reader can consult the proof in [1, p. 27]. The
proof of the second result (which can also be proved easily by verifying the four Penrose
conditions) is as follows:
(A
T
)+ = lim (AA
T
+ 8
2
l)-IA
~ - - + O
= lim [AT(AAT + 8
2
l)-1{
~ - - + O
= [limAT(AAT + 8
2
l)-1{
~ - - + O
= (A+{. 0
Note that by combining Theorems 4.12 and 4.13 we can, in theory at least, compute
the Moore-Penrose pseudoinverse of any matrix (since A A
T
and A
T
A are symmetric). This
turns out to be a poor approach in finite-precision arithmetic, however (see, e.g., [7], [11],
[23]), and better methods are suggested in text that follows.
Theorem 4.11 is suggestive of a "reverse-order" property for pseudoinverses of prod-
nets of matrices such as exists for inverses of nroducts TTnfortnnatelv. in peneraK
As an example consider A = [0 1J and B = LI. Then
while
However, necessary and sufficient conditions under which the reverse-order property does
hold are known and we quote a couple of moderately useful results for reference.
Theorem 4.14. (AB)
+
= B
+
A
+
if and only if
Proof: For the proof, see [9]. D
Theorem 4.15. (AB)
+
= B?A+, where BI = A+AB and A) = AB\B+.
Theorem 4.16. If A e R
n
r
xr
, B e R
r
r
xm
, then (AB)
+
= B+A+.
Proof: Since A e R
n
r
xr
, then A
+
= (A
T
A)~
l
A
T
, whence A
+
A = I
r
. Similarly, since
B e W
r
xm
, we have B
+
= B
T
(BB
T
)~\ whence BB
+
= I
r
. The result then follows by
taking BI = B, A\ = A in Theorem 4.15. D
The following theorem gives some additional useful properties of pseudoinverses.
Theorem 4.17. For all A e R
mxn
,
32 Chapter 4. Introduction to the Moore-Penrose Pseudo inverse
Note that by combining Theorems 4.12 and 4.13 we can, in theory at least, compute
the Moore-Penrose pseudoinverse of any matrix (since AAT and AT A are symmetric). This
turns out to be a poor approach in finite-precision arithmetic, however (see, e.g., [7], [II],
[23]), and better methods are suggested in text that follows.
Theorem 4.11 is suggestive of a "reverse-order" property for pseudoinverses of prod-
ucts of matrices such as exists for inverses of products. Unfortunately, in general,
As an example consider A = [0 I] and B = [ : J. Then
(AB)+ = 1+ = I
while
B+ A+ = [ ] =
However, necessary and sufficient conditions under which the reverse-order property does
hold are known and we quote a couple of moderately useful results for reference.
Theorem 4.14. (AB)+ = B+ A + if and only if
1. n(BB
T
AT) n(AT)
and
2. n(A T AB) nCB) .
Proof: For the proof, see [9]. 0
Theorem 4.15. (AB)+ = B{ Ai, where BI = A+ AB and AI = ABIB{.
Theorem 4.16. If A E B E then (AB)+ = B+ A+.
Proof' Since A E then A+ = (AT A)-I AT, whence A+ A = f
r
Similarly, since
B E lR;xm, we have B+ = BT(BBT)-I, whence BB+ = f
r
. The result then follows by
takingB
t
= B,At = A in Theorem 4.15. 0
The following theorem gives some additional useful properties of pseudoinverses.
Theorem 4.17. For all A E lR
mxn
,
1. (A+)+ = A.
2. (AT A)+ = A+(A
T
)+, (AA
T
)+ = (A
T
)+ A+.
3. n(A+) = n(A
T
) = n(A+ A) = n(A
T
A).
4. N(A+) = N(AA+) = NAA
T
)+) = N(AA
T
) = N(A
T
).
5. If A is normal, then AkA+ = A+ Ak and (Ak)+ = (A+)kforall integers k > O.
Exercises 33
Note: Recall that A e R"
xn
is normal if AA
T
= A
T
A. For example, if A is symmetric,
skew-symmetric, or orthogonal, then it is normal. However, a matrix can be none of the
preceding but still be normal, such as
for scalars a, b e E.
The next theorem is fundamental to facilitating a compact and unifying approach
to studying the existence of solutions of (matrix) linear equations and linear least squares
problems.
Theorem 4.18. Suppose A e R
nxp
, B e E
MX m
. Then K(B) c U(A) if and only if
AA+B = B.
Proof: Suppose K(B) c U(A) and take arbitrary jc e R
m
. Then Bx e H(B) c H(A), so
there exists a vector y e R
p
such that Ay = Bx. Then we have
where one of the Penrose properties is used above. Since x was arbitrary, we have shown
that B = AA+B.
To prove the converse, assume that AA
+
B = B and take arbitrary y e K(B). Then
there exists a vector x e R
m
such that Bx = y, whereupon
EXERCISES
1. Use Theorem 4.4 to compute the pseudoinverse of \
2 2
1
2. If jc, y e R", show that (xy
T
)
+
= (x
T
x)
+
(y
T
y)
+
yx
T
.
3. For A e R
mxn
, prove that 7(A) = 7(AA
r
) using only definitions and elementary
properties of the Moore-Penrose pseudoinverse.
4. For A e R
mxn
, prove that ft(A+) = ft(A
r
).
5. For A e R
pxn
and 5 R
mx
", show that JV(A) C A/"(S) if and only if fiA+A = B.
6. Let A G M"
xn
, 5 e E
nxm
, and D E
mxm
and suppose further that D is nonsingular.
(a) Prove or disprove that
(b) Prove or disprove that
Exercises 33
Note: Recall that A E IRn xn is normal if A A T = A T A. For example, if A is symmetric,
skew-symmetric, or orthogonal, then it is normal. However, a matrix can be none of the
preceding but still be normal, such as
A=[ a b]
-b a
for scalars a, b E R
The next theorem is fundamental to facilitating a compact and unifying approach
to studying the existence of solutions of (matrix) linear equations and linear least squares
problems.
Theorem 4.18. Suppose A E IRnxp, B E IRnxm. Then R(B) S; R(A) if and only if
AA+B = B.
Proof: Suppose R(B) S; R(A) and take arbitrary x E IRm. Then Bx E R(B) S; RCA), so
there exists a vector y E IRP such that Ay = Bx. Then we have
Bx = Ay = AA + Ay = AA + Bx,
where one of the Penrose properties is used above. Since x was arbitrary, we have shown
that B = AA+ B.
To prove the converse, assume that AA + B = B and take arbitrary y E R(B). Then
there exists a vector x E IR
m
such that Bx = y, whereupon
y = Bx = AA+Bx E R(A). 0
EXERCISES
1. Use Theorem 4.4 to compute the pseudoinverse of U ;].
2. If x, Y E IRn, show that (xyT)+ = (x
T
x)+(yT y)+ yx
T
.
3. For A E IRmxn, prove that RCA) = R(AAT) using only definitions and elementary
properties of the Moore-Penrose pseudoinverse.
4. For A E IRmxn, prove that R(A+) = R(A
T
).
5. For A E IRPxn and BE IRmxn, show thatN(A) S; N(B) if and only if BA+ A = B.
6. Let A E IRn xn, B E JRn xm , and D E IRm xm and suppose further that D is nonsingular.
(a) Prove or disprove that
[
AB
r = [
A+ -A+ABD-
i
].
D 0
D-
i
(b) Prove or disprove that
[
B
r =[
A+ -A+BD-
1
l
D 0
D-
i
Chapter 5
Introduction to the Singular
Value Decomposition
In this chapter we give a brief introduction to the singular value decomposition (SVD). We
show that every matrix has an SVD and describe some useful properties and applications
of this important matrix factorization. The SVD plays a key conceptual and computational
role throughout (numerical) linear algebra and its applications.
5.1 The Fundamental Theorem
xn
. Then there exist orthogonal matrices U e R
mxm
and
V R
nxn
such that
where S = [J
0
], S = diagfcri, ... , o>) e R
rxr
, and a\ > > o
r
> 0. More
specifically, we have
The submatrix sizes are all determined by r (which must be < min{m, }), i.e., U\ e W
nxr
,
U
2 e
^x(m-r)
; Vi e R
xr
j
y
2
Rnxfo-r^
and the
0-
JM
^/ocJb in E are compatibly
dimensioned.
Proof: Since A
r
A> 0 ( A
r
Ai s symmetric and nonnegative definite; recall, for example,
[24, Ch. 6]), its eigenvalues are all real and nonnegative. (Note: The rest of the proof follows
analogously if we start with the observation that A A
T
> 0 and the details are left to the reader
as an exercise.) Denote the set of eigenvalues of A
T
A by {of , / e n} with a\ > > a
r
>
0 = o>
+
i = = a
n
. Let {u, , i e n} be a set of corresponding orthonormal eigenvectors
and let V\ = [v\, ..., v
r
] , Vi = [v
r+
\, . . . , v
n
]. Letting S diag(cri, . . . , cf
r
), we can
write A
r
AVi = ViS
2
. Premultiplying by Vf gives Vf A
T
AVi = VfV^S
2
= S
2
, the latter
equality following from the orthonormality of the r;, vectors. Pre- and postmultiplying by
S~
l
eives the emotion
35
Chapter 5
Introduction to the Singular
Value Decomposition
In this chapter we give a brief introduction to the singular value decomposition (SVD). We
show that every matrix has an SVD and describe some useful properties and applications
of this important matrix factorization. The SVD plays a key conceptual and computational
role throughout (numerical) linear algebra and its applications.
5.1 The Fundamental Theorem
Theorem 5.1. Let A E Then there exist orthogonal matrices U E IRmxm and
V E IR
nxn
such that
A =
(5.1)
where =
n
S diag(ul, ... , u
r
) E
IRrxr, and UI
> > U
r
> O. More
specifically, we have
U2) [
0
] [
V
T
]
A = [U
I
I
(5.2)
0
VT
2
= Ulsvt
(5.3)
The submatrix sizes are all determined by r (which must be S min{m, n}), i.e., UI E IRmxr,
U2 E IRrnx(m-rl, VI E IRnxr, V
2
E IRnx(n-r), and the O-subblocks in are compatibly
dimensioned.
Proof: Since AT A ::::: 0 (AT A is symmetric and nonnegative definite; recall, for example,
[24, Ch. 6]), its eigenvalues are all real and nonnegative. (Note: The rest of the proof follows
analogously if we start with the observation that AAT ::::: 0 and the details are left to the reader
as an exercise.) Denote the set of eigenvalues of AT A by {U?, i E !!.} with UI ::::: ... ::::: U
r
>
0= Ur+1 = ... = Un. Let {Vi, i E !!.} be a set of corresponding orthonormal eigenvectors
and let VI = [VI, ... ,V
r
), V2 = [Vr+I, ... ,V
n
]. LettingS = diag(uI, ... ,u
r
), we can
write A T A VI = VI S2. Premultiplying by vt gives vt A T A VI = vt VI S2 = S2, the latter
equality following from the orthonormality of the Vi vectors. Pre- and postmultiplying by
S-I gives the equation
(5.4)
35
36 Chapter 5. Introduction to the Singular Value Decomposition
Turning now to the eigenvalue equations corresponding to the eigenvalues o
r+
\, . . . , a
n
we
have that A
T
AV
2
= V
2
0 = 0, whence Vf A
T
AV
2
= 0. Thus, AV
2
= 0. Now define the
matrix Ui e M
mx/
" by U\ = AViS~
l
. Then from (5.4) we see that UfU\ = /; i.e., the
columns of U\ are orthonormal. Choose any matrix U
2
^
7 7 I X(
~
r)
such that [U\ U
2
] is
orthogonal. Then
since A V
2
=0. Referring to the equation U\ = A V\ S
l
defining U\, we see that U{ AV\ =
S and 1/2 AVi = UÛiS = 0. The latter equality follows from the orthogonality of the
columns of U\ andU
2
. Thus, we see that, in fact, U
T
AV = [Q Q], and defining this matrix
to be S completes the proof. D
Definition 5.2. Let A = t/E V
T
be an SVD of A as in Theorem 5.1.
1. The set [a\,..., a
r
} is called the set of (nonzero) singular values of the matrix A and
i
is denoted (A). From the proof of Theorem 5.1 we see that cr,(A) = A
(
2
(A
T
A) =
A.? (AA
T
). Note that there are alsomin{m, n] r zero singular values.
2. The columns ofU are called the left singular vectors of A (and are the orthonormal
eigenvectors of AA
T
).
3. The columns of V are called the right singular vectors of A (and are the orthonormal
eigenvectors of A
1
A).
Remark 5.3. The analogous complex case in which A e C
x
" is quite straightforward.
The decomposition is A = t/E V
H
, where U and V are unitary and the proof is essentially
identical, except for Hermitian transposes replacing transposes.
Remark 5.4. Note that U and V can be interpreted as changes of basis in both the domain
and co-domain spaces with respect to which A then has a diagonal matrix representation.
Specifically, let C, denote A thought of as a linear transformation mapping W to W. Then
rewriting A = U^V
T
as AV = U E we see that Mat is S with respect to the bases
[v\,..., v
n
} for R" and {u\,..., u
m
] for R
m
(see the discussion in Section 3.2). See also
Remark 5.16.
Remark 5.5. The singular value decomposition is not unique. For example, an examination
of the proof of Theorem 5.1 reveals that
any orthonormal basis for jV(A) can be used for V
2
.
there may be nonuniqueness associated with the columns of V\ (and hence U\) cor-
responding to multiple cr/' s.
Turning now to the eigenvalue equations corresponding to the eigenvalues ar+l, ... , an we
have that A T A V
z
= VzO = 0, whence Vi A T A V
2
= O. Thus, A V
2
= O. Now define the
matrix VI E IRmxr by VI = AViS-I. Then from (5.4) we see that VrVI = /; i.e., the
columns of VI are orthonormal. Choose any matrix V2 E IRmx(m-r) such that [VI V2] is
orthogonal. Then
V
T
AV = [
VrAV
I
VIAV
I
=[
VrAV
I
vIA VI
Vr AV
z
]
vI AV
z
]
since A V
2
= O. Referring to the equation V I = A VI S-I defining VI, we see that V r A VI =
S and vI A VI = vI VI S = O. The latter equality follows from the orthogonality of the
columns of VI and V
2
. Thus, we see that, in fact, VT A V = and defining this matrix
to be completes the proof. 0
Definition 5.2. Let A = V"i:. VT be an SVD of A as in Theorem 5.1.
1. The set {ai, ... , a
r
} is called the set of (nonzero) singular values of the matrix A and
I
is denoted From the proof of Theorem 5.1 we see that ai(A) = A;'- (AT A) =
I
At- (AA
T
). Note that there are also min{m, n} - r zero singular values.
2. The columns of V are called the left singular vectors of A (and are the orthonormal
eigenvectors of AA
T
).
3. The columns of V are called the right singular vectors of A (and are the orthonormal
eigenvectors of AT A).
Remark 5.3. The analogous complex case in which A E xn is quite straightforward.
The decomposition is A = V"i:. V H, where V and V are unitary and the proof is essentially
identical, except for Hermitian transposes replacing transposes.
Remark 5.4. Note that V and V can be interpreted as changes of basis in both the domain
and co-domain spaces with respect to which A then has a diagonal matrix representation.
Specifically, let C denote A thought of as a linear transformation mapping IR
n
to IRm. Then
rewriting A = V"i:. VT as A V = V"i:. we see that Mat C is "i:. with respect to the bases
{VI, ... , v
n
} for IR
n
and {u I, .. , u
m
} for IR
m
(see the discussion in Section 3.2). See also
Remark 5.16.
Remark 5.5. The !:ingular value decomposition is not unique. For example, an examination
of the proof of Theorem 5.1 reveals that
lny orthonormal basis for N(A) can be used for V2.
there may be nonuniqueness associated with the columns of VI (and hence VI) cor-
responding to multiple O'i'S.
5.1. The Fundamental Theorem 37
any C/
2
can be used so long as [U\ Ui] is orthogonal.
columns of U and V can be changed (in tandem) by sign (or multiplier of the form
e
je
in the complex case).
What is unique, however, is the matrix E and the span of the columns of U\, f/2, Vi, and
2 (see Theorem 5.11). Note, too, that a "full SVD" (5.2) can always be constructed from
a "compact SVD" (5.3).
Remark 5.6. Computing an SVD by working directly with the eigenproblem for A
T
A or
AA
T
is numerically poor in finite-precision arithmetic. Better algorithms exist that work
directly on A via a sequence of orthogonal transformations; see, e.g., [7], [11], [25].
F/vamnlp 5.7.
Example 5.10. Let A e R
MX
" be symmetric and positive definite. Let V be an orthogonal
matrix of eigenvectors that diagonalizes A, i.e., V
T
AV = A > 0. Then A = VAV
T
is an
SVD of A.
A factorization t/SV
r
o f a n m x n matrix A qualifies as an SVD if U and V are
orthogonal and is an m x n "diagonal" matrix whose diagonal elements in the upper
left corner are positive (and ordered). For example, if A = f/E V
T
is an SVD of A, then
VS
r
C/
r
i sanSVDof A
T
.
where U is an arbitrary 2x2 orthogonal matrix, is an SVD.
Example 5.8.
where 0 is arbitrary, is an SVD.
Example 5.9.
is an SVD.
5.1. The Fundamental Theorem 37
any U2 can be used so long as [U
I
U2] is orthogonal.
columns of U and V can be changed (in tandem) by sign (or multiplier of the form
e
j8
in the complex case).
What is unique, however, is the matrix I: and the span of the columns of UI, U2, VI, and
V
2
(see Theorem 5.11). Note, too, that a "full SVD" (5.2) can always be constructed from
a "compact SVD" (5.3).
Remark 5.6. Computing an SVD by working directly with the eigenproblem for A T A or
AA T is numerically poor in finite-precision arithmetic. Better algorithms exist that work
directly on A via a sequence of orthogonal transformations; see, e.g., [7], [11], [25],
Example 5.7.
A - [1 0 ] - U I U
T
- 01- ,
where U is an arbitrary 2 x 2 orthogonal matrix, is an SVD.
Example 5.8.
A _ [ 1
- 0 - ~ ] = [
where e is arbitrary, is an SVD.
Example 5.9.
cose
- sine
sin e
cose J [ ~ ~ J [
cose
sine
A=U n=[
I -2y'5
2 ~ ][ ~ 0][
3
-5-
2
y'5
4y'5 0 0
3 S- 15
2
0
_y'5 0 0
3
-3-
[
I
]
3
3J2 [ ~
~ ]
=
2
3
2
3
is an SVD.
Sine]
-cose '
v'2 v'2
]
T T
v'2 -v'2
T
-2-
Example 5.10. Let A E IR
nxn
be symmetric and positive definite. Let V be an orthogonal
matrix of eigenvectors that diagonalizes A, i.e., VT A V = A > O. Then A = V A V
T
is an
SVDof A.
A factorization UI: VT of an m x n matrix A qualifies as an SVD if U and V are
orthogonal and I: is an m x n "diagonal" matrix whose diagonal elements in the upper
left comer are positive (and ordered). For example, if A = UI:V
T
is an SVD of A, then
VI:TU
T
is an SVD of AT.
5.2 Some Basic Properties
mxn
have a singular value decomposition A = VLV
T
. Using
the notation of Theorem 5.1, the following properties hold:
1. rank(A) = r = the number of nonzero singular values of A.
2. Let U =. [ H I , ..., u
m
] and V = [v\, ..., v
n
]. Then A has the dyadic (or outer
product) expansion
Remark 5.12. Part 4 of the above theorem provides a numerically superior method for
finding (orthonormal) bases for the four fundamental subspaces compared to methods based
on, for example, reduction to row or column echelon form. Note that each subspace requires
knowledge of the rank r. The relationship to the four fundamental subspaces is summarized
nicely in Figure 5.1.
Remark 5.13. The elegance of the dyadic decomposition (5.5) as a sum of outer products
and the key vector relations (5.6) and (5.7) explain why it is conventional to write the SVD
as A = UZV
T
rather than, say, A = UZV.
Theorem 5.14. Let A e E
mx
" have a singular value decomposition A = UHV
T
as in
TheoremS.]. Then
where
3. The singular vectors satisfy the relations
5.2 Some Basic Properties
Theorem 5.11. Let A E jRrnxn have a singular value decomposition A = U' V
T
. Using
the notation of Theorem 5.1, the following properties hold:
1. rank(A) = r = the number of nonzero singular values of A.
2. Let V = [UI, ... , urn] and V = [VI, ... , v
n
]. Then A has the dyadic (or outer
product) expansion
r
A = Laiuiv;.
i=1
3. The singular vectors satisfy the relations
for i E r.
AVi = ajui,
AT Uj = aivi
(5.5)
(5.6)
(5.7)
4. LetUI = [UI, ... , u
r
], U2 = [Ur+I, ... , urn], VI = [VI, ... , v
r
], andV2 = [Vr+I, ... , V
n
].
Then
(a) R(VI) = R(A) = N(A
T
/.
(b) R(U
2
) = R(A)1- = N(A
T
).
(c) R(VI) = N(A)1- = R(A
T
).
(d) R(V2) = N(A) = R(AT)1-.
Remark 5.12. Part 4 of the above theorem provides a numerically superior method for
finding (orthonormal) bases for the four fundamental subspaces compared to methods based
on, for example, reduction to row or column echelon form. Note that each subspace requires
knowledge of the rank r. The relationship to the four fundamental subspaces is summarized
nicely in Figure 5.1.
Remark 5.13. The elegance of the dyadic decomposition (5.5) as a sum of outer products
and the key vector relations (5.6) and (5.7) explain why it is conventional to write the SVD
as A = U'V
T
rather than, say, A = U,V.
Theorem 5.14. Let A E jRmxn have a singular value decomposition A = U,V
T
as in
Theorem 5.1. Then
(5.8)
where
5.2. Some Basic Properties 39
Figure 5.1. SVD and the four fundamental subspaces.
with the Q-subblocks appropriately sized. Furthermore, if we let the columns of U and V
be as defined in Theorem 5.11, then
Proof: The proof follows easily by verifying the four Penrose conditions. D
Remark 5.15. Note that none of the expressions above quite qualifies as an SVD of A
+
if we insist that the singular values be ordered from largest to smallest. However, a simple
reordering accomplishes the task:
This can also be written in matrix terms by using the so-called reverse-order identity matrix
(or exchange matrix) P = \e
r
,e
r
^\, ..., e^, e\\, which is clearly orthogonal and symmetric.
5.2. Some Basic Properties 39
A
r r
E9 {O}
/ {O)<!l
n-r m-r
Figure 5.1. SVD and the four fundamental subspaces.
with the O-subblocks appropriately sized. Furthermore, if we let the columns of U and V
be as defined in Theorem 5.11, then
r 1
= L -v;u;, (5.10)
;=1 U;
Proof' The proof follows easily by verifying the four Penrose conditions. 0
Remark 5.15. Note that none of the expressions above quite qualifies as an SVD of A +
if we insist that the singular values be ordered from largest to smallest. However, a simple
reordering accomplishes the task:
(5.11)
This can also be written in matrix terms by using the so-called reverse-order identity matrix
(or exchange matrix) P = [e
r
, er-I, ... , e2, ed, which is clearly orthogonal and symmetric.
is the matrix version of (5.11). A "full SVD" can be similarly constructed.
Remark 5.16. Recall the linear transformation T used in the proof of Theorem 3.17 and
in Definition 4.1. Since T is determined by its action on a basis, and since ( v \ , . . . , v
r
} is a
basis forJ\f(A)
, then T can be defined by TV; = cr, w, , / e r. Similarly, since [u\, ... ,u

r
}
isabasisfor7(.4), then T~
l
can be defined by T^' M, = ^-u, , / e r. From Section 3.2, the
matrix representation for T with respect to the bases { v \ , ..., v
r
} and { MI , . . . , u
r
] is clearly
S, while the matrix representation for the inverse linear transformation T~
l
with respect to
the same bases is 5""
1
.
5.3 Row and Column Compressions
Row compression
Let A E R
mxn
have an SVD given by (5.1). Then
Notice that M(A) - M(U
T
A) = A/"(SV,
r
) and the matrix SVf e R
r x
" has full row
rank. I n other words, premultiplication of A by U
T
is an orthogonal transformation that
"compresses" A by row transformations. Such a row compression can also be accomplished
D _
by orthogonal row transformations performed directly on A to reduce it to the form
0
,
where R is upper triangular. Both compressions are analogous to the so-called row-reduced
echelon form which, when derived by a Gaussian elimination algorithm implemented in
finite-precision arithmetic, is not generally as reliable a procedure.
Column compression
Again, let A e R
mxn
This time, notice that H(A) = K(AV) = K(UiS) and the matrix UiS e R
mxr
has full
column rank. I n other words, postmultiplication of A by V is an orthogonal transformation
that "compresses" A by column transformations. Such a compression is analogous to the
40 Chapters. Introduction to the Singular Value Decomposition
Then
Then
A+ = (VI p)(PS-1 p)(PVr)
is the matrix version of (5.11). A "full SVD" can be similarly constructed.
Remark 5.16. Recall the linear transformation T used in the proof of Theorem 3.17 and
in Definition 4.1. Since T is determined by its action on a basis, and since {VI, ... , v
r
} is a
basisforN(A).l, then T can be defined by TVj = OjUj , i E ~ . Similarly, since {UI, ... , u
r
}
is a basis forR(A), then T-
I
canbedefinedbyT-Iu; = tv; ,i E ~ . From Section 3.2, the
matrix representation for T with respect to the bases {VI, ... , v
r
} and {u I, ... , u
r
} is clearly
S, while the matrix representation for the inverse linear transformation T-
I
with respect to
the same bases is S-I.
5.3 Rowand Column Compressions
Row compression
Let A E lR.
mxn
VT A = :EVT
= [ ~ ~ ] [ ~ i ]
- [ SVr ] lR.
mxn
- 0 E .
Notice that N(A) = N(V
T
A) = N(svr> and the matrix SVr E lR.
rxll
has full row
rank. In other words, premultiplication of A by VT is an orthogonal transformation that
"compresses" A by row transformations. Such a row compression can also be accomplished
by orthogonal row transformations performed directly on A to reduce it to the form [ ~ ] ,
where R is upper triangular. Both compressions are analogous to the so-called row-reduced
echelon form which, when derived by a Gaussian elimination algorithm implemented in
finite-precision arithmetic, is not generally as reliable a procedure.
Column compression
Again, let A E lR.
mxn
AV = V:E
= [VI U2] [ ~ ~ ]
=[VIS 0] ElR.mxn.
This time, notice that R(A) = R(A V) = R(UI S) and the matrix VI S E lR.
m
xr has full
column rank. In other words, postmultiplication of A by V is an orthogonal transformation
that "compresses" A by I;olumn transformations. Such a compression is analogous to the
Exercises 41
so-called column-reduced echelon form, which is not generally a reliable procedure when
performed by Gauss transformations in finite-precision arithmetic. For details, see, for
example, [7], [11], [23], [25].
EXERCISES
1. Let X M
mx
". If X
T
X = 0, show that X = 0.
2. Prove Theorem 5.1 starting from the observation that AA
T
> 0.
3. Let A e E"
xn
be symmetric but indefinite. Determine an SVD of A.
4. Let x e R
m
, y e R
n
be nonzero vectors. Determine an SVD of the matrix A e R
defined by A = xy
T
.
6. Let A e R
mxn
and suppose W eR
mxm
and 7 e R
nxn
are orthogonal.
(a) Show that A and W A F have the same singular values (and hence the same rank).
(b) Suppose that W and Y are nonsingular but not necessarily orthogonal. Do A
and WAY have the same singular values? Do they have the same rank?
7. Let A R"
XM
. Use the SVD to determine a polar factorization of A, i.e., A = QP
where Q is orthogonal and P = P
T
> 0. Note: this is analogous to the polar form
z = re
l&
ofa complex scalar z (where i = j = V^T).
5. Determine SVDs of the matrices
Exercises 41
so-called column-reduced echelon form, which is not generally a reliable procedure when
performed by Gauss transformations in finite-precision arithmetic. For details, see, for
example, [7], [11], [23], [25].
EXERCISES
1. Let X E IRmxn. If XT X = 0, show that X = o.
2. Prove Theorem 5.1 starting from the observation that AAT ~ O.
3. Let A E IR
nxn
be symmetric but indefinite. Determine an SVD of A.
4. Let x E IRm, y E ~ n be nonzero vectors. Determine an SVD of the matrix A E ~ ~ xn
defined by A = xyT.
5. Determine SVDs of the matrices
(a)
[
-1
]
0 -1
(b)
[
~ l
6. Let A E ~ m x n and suppose W E IR
mxm
and Y E ~ n x n are orthogonal.
(a) Show that A and WAY have the same singular values (and hence the same rank).
(b) Suppose that Wand Yare nonsingular but not necessarily orthogonal. Do A
and WAY have the same singular values? Do they have the same rank?
7. Let A E ~ ~ x n . Use the SVD to determine a polar factorization of A, i.e., A = Q P
where Q is orthogonal and P = p
T
> O. Note: this is analogous to the polar form
z = re
iO
of a complex scalar z (where i = j = J=I).
Chapter 6
Li near Equations
In this chapter we examine existence and uniqueness of solutions of systems of linear
equations. General linear systems of the form
are studied and include, as a special case, the familiar vector system
6.1 Vector Li near Equations
We begin with a review of some of the principal results associated with vector linear systems.
Theorem 6.1. Consider the system of linear equations
1. There exists a solution to (6.3) if and only ifbeH(A).
2. There exists a solution to (6.3} for all b e R
m
if and only ifU(A) = W", i.e., A is
onto; equivalently, there exists a solution if and only j/"rank([A, b]) = rank(A), and
this is possible only ifm < n (since m = dimT^(A) = rank(A) < min{m, n}).
3. A solution to (6.3) is unique if and only ifJ\f(A) = 0, i.e., A is 1-1.
4. There exists a unique solution to (6.3) for all b e W" if and only if A is nonsingular;
equivalently, A G M
mxm
and A has neither a 0 singular value nor a 0 eigenvalue.
5. There exists at most one solution to (6.3) for all b e W
1
if and only if the columns of
A are linearly independent, i.e., A/"(A) = 0, and this is possible only ifm > n.
6. There exists a nontrivial solution to the homogeneous system Ax = 0 if and only if
rank(A) < n.
43
Chapter 6
Linear Equations
In this chapter we examine existence and uniqueness of solutions of systems of linear
equations. General linear systems of the form
(6.1)
are studied and include, as a special case, the familiar vector system
Ax = b; A E ]Rn xn, b E ]Rn.
(6.2)
6.1 Vector Linear Equations
We begin with a review of some of the principal results associated with vector linear systems.
Theorem 6.1. Consider the system of linear equations
Ax = b; A E lR
m
xn, b E lRm.
(6.3)
1. There exists a solution to (6.3) if and only if b E R(A).
2. There exists a solution to (6.3) for all b E lR
m
if and only ifR(A) = lR
m
, i.e., A is
onto; equivalently, there exists a solution if and only ifrank([A, b]) = rank(A), and
this is possible only ifm :::: n (since m = dim R(A) = rank(A) :::: min{m, n n.
3. A solution to (6.3) is unique if and only if N(A) = 0, i.e., A is 1-1.
4. There exists a unique solution to (6.3) for all b E ]Rm if and only if A is nonsingular;
equivalently, A E lR
mxm
and A has neither a 0 singular value nor a 0 eigenvalue.
5. There exists at most one solution to (6.3) for all b E lR
m
if and only if the columns of
A are linearly independent, i.e., N(A) = 0, and this is possible only ifm ::: n.
6. There exists a nontrivial solution to the homogeneous system Ax = 0 if and only if
rank(A) < n.
43
44 Chapter 6. Linear Equations
Proof: The proofs are straightforward and can be consulted in standard texts on linear
algebra. Note that some parts of the theorem follow directly from others. For example, to
prove part 6, note that x = 0 is always a solution to the homogeneous system. Therefore, we
must have the case of a nonunique solution, i.e., A is not 1-1, which implies rank(A) < n
by part 3. D
6.2 Matrix Linear Equations
In this section we present some of the principal results concerning existence and uniqueness
of solutions to the general matrix linear system (6.1). Note that the results of Theorem
6.1 follow from those below for the special case k = 1, while results for (6.2) follow by
specializing even further to the case m = n.
Theorem 6.2 (Existence). The matrix linear equation
and this is clearly of the form (6.5).
has a solution if and only ifl^(B) C 7(A); equivalently, a solution exists if and only if
AA
+
B = B.
Proof: The subspace inclusion criterion follows essentially from the definition of the range
of a matrix. The matrix criterion is Theorem 4.18.
mxn
, B eR
mxk
and suppose that AA
+
B = B. Then any matrix
of the form
is a solution of
Furthermore, all solutions of (6.6) are of this form.
Proof: To verify that (6.5) is a solution, premultiply by A:
That all solutions arc of this form can be seen as follows. Let Z be an arbitrary solution of
(6.6), i.e., AZ B. Then we can write
Proof: The proofs are straightforward and can be consulted in standard texts on linear
algebra. Note that some parts of the theorem follow directly from others. For example, to
prove part 6, note that x = 0 is always a solution to the homogeneous system. Therefore, we
must have the case of a nonunique solution, i.e., A is not I-I, which implies rank(A) < n
by part 3. 0
6.2 Matrix Linear Equations
In this section we present some of the principal results concerning existence and uniqueness
of solutions to the general matrix linear system (6.1). Note that the results of Theorem
6.1 follow from those below for the special case k = 1, while results for (6.2) follow by
specializing even further to the case m = n.
Theorem 6.2 (Existence). The matrix linear equation
AX = B; A E JR.
mxn
, BE JR.mxk, (6.4)
has a solution if and only ifR(B) S; R(A); equivalently, a solution exists if and only if
AA+B = B.
Proof: The subspace inclusion criterion follows essentially from the definition of the range
of a matrix. The matrix criterion is Theorem 4.18. 0
Theorem 6.3. Let A E JR.mxn, B E JR.mxk and suppose that AA + B = B. Then any matrix
of the form
X = A+ B + (/ - A+ A)Y, where Y E JR.nxk is arbitrary, (6.5)
is a solution of
AX=B. (6.6)
Furthermore, all solutions of (6.6) are of this form.
Proof: To verify that (6.5) is a solution, premultiply by A:
AX = AA+ B + A(I - A+ A)Y
= B + (A - AA+ A)Y by hypothesis
= B since AA + A = A by the first Penrose condition.
That all solutions are of this form can be seen as follows. Let Z be an arbitrary solution of
(6.6). i.e .. AZ :::: B. Then we can write
Z=A+AZ+(I-A+A)Z
=A+B+(I-A+A)Z
and this is clearly of the form (6.5). 0
6.2. Matrix Linear Equations 45
Remark 6.4. When A is square and nonsingular, A
+
= A"
1
and so (/ A
+
A) = 0. Thus,
there is no "arbitrary" component, leaving only the unique solution X = A~
1
B.
Remark 6.5. It can be shown that the particular solution X = A
+
B is the solution of (6.6)
that minimizes TrX
7
X. (Tr(-) denotes the trace of a matrix; recall that TrX
r
X = \ jcj.)
Theorem 6.6 (Uniqueness). A solution of the matrix linear equation
is unique if and only if A
+
A = /; equivalently, (6.7) has a unique solution if and only if
M(A) = 0.
Proof: The first equivalence is immediate from Theorem 6.3. The second follows by noting
that A
+
A = / can occur only if r n, where r = rank(A) (recall r < h). But rank(A) = n
if and only if A is 1-1 or _ /V(A) = 0. D
Example 6.7. Suppose A e E"
x
". Find all solutions of the homogeneous system Ax 0.
Solution:
where y e R" is arbitrary. Hence, there exists a nonzero solution if and only if A
+
A /= I.
This is equivalent to either rank (A) = r < n or A being singular. Clearly, if there exists a
nonzero solution, it is not unique.
Computation: Since y is arbitrary, it is easy to see that all solutions are generated
from a basis for 7(7 A
+
A). But if A has an SVD given by A = f/E V
T
, then it is easily
checked that / - A+A = V
2
V
2
r
and U(V
2
V^) = K(V
2
) = N(A).
Example 6.8. Characterize all right inverses of a matrix A e ]R
mx
"; equivalently, find all
solutions R of the equation AR = I
m
. Here, we write I
m
to emphasize the m x m identity
matrix.
Solution: There exists a right inverse if and only if 7(/
m
) c 7(A) and this is
equivalent to AA
+
I
m
= I
m
. Clearly, this can occur if and only if rank(A) = r = m (since
r < m) and this is equivalent to A being onto (A
+
is then a right inverse). All right inverses
of A are then of the form
where Y e E"
xm
is arbitrary. There is a unique right inverse if and only if A
+
A = /
(AA(A) = 0), in which case A must be invertible and R = A"
1
.
Example 6.9. Consider the system of linear first-order difference equations
6.2. Matrix Linear Equations 45
Remark 6.4. When A is square and nonsingular, A + = A-I and so (I - A + A) = O. Thus,
there is no "arbitrary" component, leaving only the unique solution X = A-I B.
Remark 6.5. It can be shown that the particular solution X = A + B is the solution of (6.6)
that minimizes TrXT X. (TrO denotes the trace of a matrix; recall that TrXT X = Li,j xlj.)
Theorem 6.6 (Uniqueness). A solution of the matrix linear equation
AX = B; A E lR,mxn, BE lR,mxk
(6,7)
is unique if and only if A + A = I; equivalently, (6.7) has a unique solution if and only if
N(A) = O.
Proof: The first equivalence is immediate from Theorem 6.3. The second follows by noting
thatA+ A = I can occur only ifr = n, wherer = rank(A) (recallr ::: n), Butrank(A) = n
if and only if A is I-lor N(A) = O. 0
Example 6.7. Suppose A E lR,nxn. Find all solutions of the homogeneous system Ax = 0,
Solution:
x=A+O+(I-A+A)y
= (I-A+A)y,
where y E lR,n is arbitrary. Hence, there exists a nonzero solution if and only if A + A t= I,
This is equivalent to either rank(A) = r < n or A being singular. Clearly, if there exists a
nonzero solution, it is not unique,
Computation: Since y is arbitrary, it is easy to see that all solutions are generated
from a basis for R(I - A + A). But if A has an SVD given by A = U h VT, then it is easily
checked that 1- A+ A = Vz V[ and R(Vz vD = R(Vz) = N(A),
Example 6.S. Characterize all right inverses of a matrix A E lR,mxn; equivalently, find all
solutions R of the equation AR = 1
m
, Here, we write 1m to emphasize the m x m identity
matrix,
Solution: There exists a right inverse if and only if R(Im) S; R(A) and this is
equivalent to AA + 1m = 1m. Clearly, this can occur if and only if rank(A) = r = m (since
r ::: m) and this is equivalent to A being onto (A + is then a right inverse). All right inverses
of A are then of the form
R = A+ 1m + (In - A+ A)Y
=A++(I-A+A)Y,
where Y E lR,nxm is arbitrary, There is a unique right inverse if and only if A+ A I
(N(A) = 0), in which case A must be invertible and R = A-I.
Example 6.9. Consider the system of linear first-order difference equations
(6,8)
with A e R"
xn
and fieR"
xm
(rc>l,ra>l). The vector Jt* in linear system theory is
known as the state vector at time k while Uk is the input (control) vector. The general
solution of (6.8) is given by
for k > 1. We might now ask the question: Given X Q = 0, does there exist an input sequence
{uj } y~ Q such that x^ takes an arbitrary va
of reachability. Since m > 1, from the
see that (6.8) is reachable if and only if
[ Uj }
k
jj^ such that X k takes an arbitrary value in W ? In linear system theory, this is a question
of reachability. Since m > 1, from the fundamental Existence Theorem, Theorem 6.2, we
or, equivalently, if and only if
A related question is the following: Given an arbitrary initial vector X Q , does there ex-
ist an input sequence {"y} "~ o such that x
n
= 0? In linear system theory, this is called
controllability. Again from Theorem 6.2, we see that (6.8) is controllable if and only if
Clearly, reachability always implies controllability and, if A is nonsingular, control-
lability and reachability are equivalent. The matrices A = [
1
Q
1 and 5 = f ^ 1 provide an
example of a system that is controllable but not reachable.
The above are standard conditions with analogues for continuous-time models (i.e.,
linear differential equations). There are many other algebraically equivalent conditions.
Example 6.10. We now introduce an output vector y
k
to the system (6.8) of Example 6.9
by appending the equation
with C e R
pxn
and D R
pxm
(p > 1). We can then pose some new questions about the
overall system that are dual in the system-theoretic sense to reachability and controllability.
The answers are cast in terms that are dual in the linear algebra sense as well. The condition
dual to reachability is called observability: When does knowledge of {"
7
}"!Q and {y_ / } "~ o
suffice to determine (uniquely) Jt
0
? As a dual to controllability, we have the notion of
reconstructibility: When does knowledge of {w
y
} "~ Q and {;y/ } "Io suffice to determine
(uniquely) x
n
l The fundamental duality result from linear system theory is the following:
(A, B) is reachable [ controllable] if and only if (A
T
, B
T
] is observable [ reconstructive].
with A E IR
nx
" and B E IR
nxm
(n I, m I). The vector Xk in linear system theory is
known as the state vector at time k while Uk is the input (control) vector. The general
solution of (6.8) is given by
k-J
Xk = Akxo + LAk-J-j BUj
j=O
k k-J Uk-2
[
Uk-J ]
... A B]
(6.9)
(6.10)
for k 1. We might now ask the question: Given Xo = 0, does there exist an input sequence
{u j 1 such that Xk takes an arbitrary value in 1R"? In linear system theory, this is a question
of reacbability. Since m I, from the fundamental Existence Theorem, Theorem 6.2, we
see that (6.8) is reachable if and only if
R([ B, AB, ... , A
n
-
J
B]) = 1R"
rank [B, AB, ... , A
n
-
J
B] = n.
A related question is the following: Given an arbitrary initial vector Xo, does there ex-
ist an input sequence {u j l'/:b such that Xn = O? In linear system theory, this is called
controllability. Again from Theorem 6.2, we see that (6.8) is controllable if and only if
Clearly, reachability always implies controllability and, if A is nonsingular, control-
lability and reachability are equivalent. The matrices A = and B = provide an
example of a system that is controllable but not reachable.
The above are standard conditions with analogues for continuous-time models (i.e.,
linear differential equations). There are many other algebraically equivalent conditions.
Example 6.10. We now introduce an output vector Yk to the system (6.8) of Example 6.9
by appending the equation
(6.11)
with C E IR
Pxn
and D E IR
Pxm
(p 1). We can then pose some new questions about the
overall system that are dual in the system-theoretic sense to reachability and controllability.
The answers are cast in terms that are dual in the linear algebra sense as well. The condition
dual to reachability is called observability: When does knowledge of {u j r/:b and {Yj l';:b
suffice to determine (uniquely) xo? As a dual to controllability, we have the notion of
reconstructibility: When does knowledge of {u j r/:b and {YJ lj:b suffice to determine
(uniquely) xn? The fundamental duality result from linear system theory is the following:
(A. B) iJ reachable [controllablcl if and only if (A T. B T) is observable [reconsrrucrible]
6.4 Some Us ef u l and I nt er es t i ng Inverses 47
To derive a condition for observability, notice that
Thus,
Let v denote the (known) vector on the left-hand side of (6.13) and let R denote the matrix on
the right-hand side. Then, by definition, v e Tl(R), so a solution exists. By the fundamental
Uniqueness Theorem, Theorem 6.6, the solution is then unique if and only if N(R) = 0,
mxn
, B e R
mxq
, and C e R
pxti
. Then the equation
has a solution if and only if AA
+
BC
+
C = B, in which case the general solution is of the
where Y R
n
*
p
is arbitrary.
A compact matrix criterion for uniqueness of solutions to (6.14) requires the notion
of the Kronecker product of matrices for its statement. Such a criterion (CC
+
< g) A
+
A I)
is stated and proved in Theorem 13.27.
6.4 Some Useful and Interesting Inverses
In many applications, the coefficient matrices of interest are square and nonsingular. Listed
below is a small collection of useful matrix identities, particularly for block matrices, as-
sociated with matrix inverses. In these identities, A e R
nxn
, B E R
nxm
, C e R
mxn
,
and D E
mxm
. Invertibility is assumed for any component or subblock whose inverse is
indicated. Verification of each identity is recommended as an exercise for the reader.
Thus,
To derive a condition for observability, notice that
k-l
Yk = CAkxo + L CAk-1-j BUj + DUk.
j=O
r
Yo - Duo
Yl - CBuo - Du]
Yn-] - L j : ~ CA
n
-
2
-j BUj - DUn-l
47
(6.12)
(6.13)
Let v denote the (known) vector on the left-hand side of (6.13) and let R denote the matrix on
the right-hand side. Then, by definition, v E R(R), so a solution exists. By the fundamental
Uniqueness Theorem, Theorem 6.6, the solution is then unique if and only if N(R) = 0,
Theorem 6.11. Let A E jRmxn, B E jRmx
q
, and C E jRpxq. Then the equation
AXC=B (6.14)
has a solution if and only if AA + BC+C = B, in which case the general solution is of the
form
(6.15)
where Y E jRnxp is arbitrary.
A compact matrix criterion for uniqueness of solutions to (6.14) requires the notion
of the Kronecker product of matrices for its statement. Such a criterion (C C+ A + A = I)
is stated and proved in Theorem 13.27.
In many applications, the coefficient matrices of interest are square and nonsingular. Listed
below is a small collection of useful matrix identities, particularly for block matrices, as-
sociated with matrix inverses. In these identities, A E jRnxn, B E jRnxm, C E jRmxn,
and D E jRm xm. Invertibility is assumed for any component or subblock whose inverse is
indicated. Verification of each identity is recommended as an exercise for the reader.
1. (A + BDCr
1
= A~
l
- A~
l
B(D~
l
+ CA~
l
B)~
[
CA~
l
.
This result is known as the Sherman-Morrison-Woodbury formula. It has many
applications (and is frequently "rediscovered") including, for example, formulas for
the inverse of a sum of matrices such as (A + D)"
1
or (A"
1
+ D"
1
) . It also
yields very efficient "updating" or "downdating" formulas in expressions such as
T 1
(A + JUT ) (with symmetric A e R"
x
" and ;c e E") that arise in optimization
theory.
EXERCISES
1. As in Example 6.8, characterize all left inverses of a matrix A e M
mx
".
2. Let A E
mx
", B e R
mxk
and suppose A has an SVD as in Theorem 5.1. Assuming
7Z(B) c 7(A), characterize all solutions of the matrix linear equation
Both of these matrices satisfy the matrix equation X^ = I from which it is obvious
that X~
l
= X. Note that the positions of the / and / blocks may be exchanged.
where E = (D CA B) (E is the inverse of the Schur complement of A). This
result follows easily from the block LU factorization in property 16 of Section 1.4.
where F = (A ED C) . This result follows easily from the block UL factor-
ization in property 17 of Section 1.4.
in terms of the SVD of A
1. (A + BDC)-I = A-I - A-IB(D-
I
+ CA-IB)-ICA-I.
This result is known as the Sherman-Morrison-Woodbury formula. It has many
applications (and is frequently "rediscovered") including, for example, formulas for
the inverse of a sum of matrices such as (A + D)-lor (A-I + D-I)-I. It also
yields very efficient "updating" or "downdating" formulas in expressions such as
(A + xx
T
) -I (with symmetric A E lR
nxn
and x E lRn) that arise in optimization
theory.
2. r
l
= [
3. !/ r
l
= l r
l
= 1
Both of these matrices satisfy the matrix equation X2 = / from which it is obvious
that X-I = X. Note that the positions of the / and - / blocks may be exchanged.
4. r
l
= [
-A-I BD-
I
]
D- I .
5. r
l
= 1
6. [ / +c
BC
r
l
= [!C / 1
7. r
l
= [ A-I l
where E = (D - CA-
I
B)-I (E is the inverse of the Schur complement of A). This
result follows easily from the block LU factorization in property 16 of Section 1.4.
8. r
l
= D-
I
l
where F = (A - B D-
I
C) -I. This result follows easily from the block UL factor-
ization in property 17 of Section 1.4.
EXERCISES
1. As in Example 6.8, characterize all left inverses of a matrix A E lR
m
xn .
2. Let A E lRmxn, B E lR
fflxk
and suppose A has an SVD as in Theorem 5.1. Assuming
R(B) R(A), characterize all solutions of the matrix linear equation
AX=B
in terms of the SVD of A.
Exercises 49
3. Let jc, y e E" and suppose further that X
T
y ^ 1. Show that
4. Let x, y E" and suppose further that X
T
y ^ 1. Show that
where c = 1/(1 x
T
y).
5. Let A e R"
x
" and let A"
1
have columns c\, ..., c
n
and individual elements y
;y
.
Assume that x/
(
7^ 0 for some / and j. Show that the matrix B A
l
ei e
T
: (i.e.,
A with subtracted from its (zy)th element) is singular.
Hint: Show that c
t
< = M(B).
6. As in Example 6.10, check directly that the condition for reconstructibility takes the
form
Exercises 49
3. Let x, y E IR
n
and suppose further that x T y i= 1. Show that
T -1 1 T
(/ - xy) = I - xy .
xTy -1
4. Let x, y E IR
n
and suppose further that x T y i= 1. Show that
-cxJ
C '
where C = 1/(1 - x
T
y).
5. Let A E 1 R ~ xn and let A -1 have columns Cl, ... ,C
n
and individual elements Yij.
Assume that Yji i= 0 for some i and j. Show that the matrix B = A - ~ i e;e; (i.e.,
A with yl subtracted from its (ij)th element) is singular.
l'
Hint: Show that Ci E N(B).
6. As in Example 6.10, check directly that the condition for reconstructibility takes the
form
N[
fA J N(A
n
).
CA
n
-
1
Chapter 7
Projections, Inner Product
Spaces, and Norms
7.1 Projections
Definition 7.1. Let V be a vector space with V = X 0 y. By Theorem 2.26, every v e V
has a unique decomposition v = x + y with x e X and y e y. Define PX y V > X c V
by
Figure 7.1. Oblique projections.
Theorem 7.2. Px,y is linear and P# y Px,y-
Theorem 7.3. A linear transformation P is a projection if and only if it is idempotent, i.e.,
P
2
= P. Also, P is a projection if and only if I P is a projection. Infact, Py,x I Px,y-
Proof: Suppose P is a projection, say on X along y (using the notation of Definition 7.1).
51
Px,y is called the (oblique) projection on X along 3^.
Figure 7.1 displays the projection of v on both X and 3^ in the case V =
Chapter 7
Projections, Inner Product
Spaces, and Norms
7.1 Projections
Definition 7.1. Let V be a vector space with V = X EEl Y. By Theorem 2.26, every v E V
has a unique decomposition v = x + y with x E X and y E y. Define pX,y : V ---+ X <; V
by
PX,yV = x for all v E V.
PX,y is called the (oblique) projection on X along y.
Figure 7.1 displays the projection of von both X and Y in the case V = ]R2.
y
x
Figure 7.1. Oblique projections.
Theorem 7.2. px.y is linear and pl.
y
= px.y.
Theorem 7.3. A linear transformation P is a projection if and only if it is idempotent, i.e.,
p2 = P. Also, P isaprojectionifandonlyifl -P isaprojection. Infact, Py.x = I -px.y.
Proof: Suppose P is a projection, say on X along Y (using the notation of Definition 7.1).
51
52 Chapter 7. Projections, Inner Product Spaces, and Norms
Let u e V be arbitrary. Then Pv = P(x + y) = Px = x. Moreover, P
2
v = PPv
Px = x = Pv. Thus, P
2
= P. Conversely, suppose P
2
= P. Let X = {v e V : Pv = v}
and y = {v V : Pv = 0}. It is easy to check that X and 3^ are subspaces. We now prove
that V = X 0 y. First note that tfveX, then Pv = v. If v e y, then Pv = 0. Hence
i f v X n y, then v = 0. Now let u e V be arbitrary. Then v = Pv + (I - P)v. Let
x = Pv, y = (I - P)v. Then Px = P
2
v = Pv = x so x e X, while Py = P(I - P}v =
Pv - P
2
v = 0 so y e y. Thus, V = X 0 y and the projection on X along y is P.
Essentially the same argument shows that / P is the projection on y along X. D
Definition 7.4. In the speci al case where y = X^, PX.X
L
*
s
called an orthogonal projec-
tion and we then use the notati on PX = PX,X
L
-
Theorem 7.5. P e E"
xn
i s the matri x of an orthogonal projecti on (onto K(P)} i f and only
i fP
2
= p = P
T
.
Proof: Let P be an orthogonal projection (on X, say, along X
L
} and let jc, y e R" be
arbitrary. Note that (/ - P)x = (I - PX,X^X = P
x
,
x
x by Theorem 7.3. Thus,
(/ - P)x e X
L
. Since Py e X, we have ( P y f ( I - P)x = y
T
P
T
(I - P)x = 0.
Since x and y were arbitrary, we must have P
T
(I P) = 0. Hence P
T
= P
T
P = P,
with the second equality following since P
T
P is symmetric. Conversely, suppose P is a
symmetric projection matrix and let x be arbitrary. Write x = Px + (I P)x. Then
x
T
P
T
(I - P)x = x
T
P(I - P}x = 0. Thus, since Px e U(P), then (/ - P)x 6 ft(P)
1
and P must be an orthogonal projection. D
7.1.1 The four fundamental orthogonal projections
Using the notation of Theorems 5.1 and 5.11, let A 6 R
mxn
with SVD A = UT,V
T
=
UtSVf. Then
are easily checked to be (unique) orthogonal projections onto the respective four funda-
mental subspaces,
Let v E V be arbitrary. Then Pv = P(x + y) = Px = x. Moreover, p
2
v = P Pv =
Px = x = Pv. Thus, p2 = P. Conversely, suppose p2 = P. Let X = {v E V : Pv = v}
and Y = {v E V : Pv = OJ. It is easy to check that X and Y are subspaces. We now prove
that V = X $ y. First note that if v E X, then Pv = v. If v E Y, then Pv = O. Hence
if v E X ny, then v = O. Now let v E V be arbitrary. Then v = Pv + (I - P)v. Let
x = Pv, y = (I - P)v. Then Px = p
2
v = Pv = x so x E X, while Py = P(l - P)v =
Pv - p
2
v = 0 so Y E y. Thus, V = X $ Y and the projection on X along Y is P.
Essentially the same argument shows that I - P is the projection on Y along X. 0
Definition 7.4. In the special case where Y = X1-, px.xl. is called an orthogonal projec-
tion and we then use the notation P
x
= PX.XL
Theorem 7.5. P E jRnxn is the matrix of an orthogonal projection (onto R(P)) if and only
if p2 = P = pT.
Proof: Let P be an orthogonal projection (on X, say, along X 1-) and let x, y E jR" be
arbitrary. Note that (I - P)x = (I - px.xJ.)x = PXJ..xx by Theorem 7.3. Thus,
(I - P)x E X1-. Since Py E X, we have (py)T (I - P)x = yT pT (I - P)x = O.
Since x and y were arbitrary, we must have pT (I - P) = O. Hence pT = pT P = P,
with the second equality following since pT P is symmetric. Conversely, suppose P is a
symmetric projection matrix and let x be arbitrary. Write x = P x + (I - P)x. Then
x
T
pT (I - P)x = x
T
P(l - P)x = O. Thus, since Px E R(P), then (I - P)x E R(P)1-
and P must be an orthogonal projection. 0
7.1 .1 The four fundamental orthogonal projections
Using the notation of Theorems 5.1 and 5.11, let A E jRmxII with SVD A = U!:V
T
U\SVr Then
r
PR(A)
AA+
U\U[
Lu;uT,
;=1
m
PR(A).L
1- AA+
U2
U
! LUiUT,
i=r+l
11
PN(A)
1- A+A
V2V{
L ViVf,
i=r+l
r
PN(A)J.
A+A
VIV{
LViVT
i=l
are easily checked to be (unique) orthogonal projections onto the respective four funda-
mental subspaces.
7.1. Projections 53
Example 7.6. Determine the orthogonal projection of a vector v e M" on another nonzero
vector w e R
n
.
Solution: Think of the vector w as an element of the one-dimensional subspace IZ(w).
Then the desired projection is simply
(using Example 4.8)
Moreover, the vector z that is orthogonal to w and such that v = Pv + z is given by
z = PK(
W
)V = (/ PK(W))V = v (^-^ j w. See Figure 7.2. A direct calculation shows
that z and u; are, in fact, orthogonal:
Figure 7.2. Orthogonal projection on a "line."
Example 7.7. Recall the proof of Theorem 3.11. There, {v\ , ..., Vk} was an orthornormal
basis for a subset S of W
1
. An arbitrary vector x e R" was chosen and a formula for x\
appeared rather mysteriously. The expression for x\ is simply the orthogonal projection of
x on S. Specifically,
Example 7.8. Recall the diagram of the four fundamental subspaces. The indicated direct
sum decompositions of the domain E" and co-domain R
m
are given easily as follows.
Let x e W
1
be an arbitrary vector. Then
7.1. Projections 53
Example 7.6. Determine the orthogonal projection of a vector v E IR
n
on another nonzero
vector w E IRn.
Solution: Think of the vector w as an element of the one-dimensional subspace R( w).
Then the desired projection is simply
Pn(w)v = ww+v
wwTv
(using Example 4.8)
= (WTV)
T W.
W W
Moreover, the vector z that is orthogonal to wand such that v = P v + z is given by
z = Pn(w)"' v = (l - Pn(wv = v - ( : ; ~ ) w. See Figure 7.2. A direct calculation shows
that z and ware, in fact, orthogonal:
v
z
Pv w
Figure 7.2. Orthogonal projection on a "line."
Example 7.7. Recall the proof of Theorem 3.11. There, {VI, ... , Vk} was an orthomormal
basis for a subset S of IRn. An arbitrary vector x E IR
n
was chosen and a formula for XI
appeared rather mysteriously. The expression for XI is simply the orthogonal projection of
X on S. Specifically,
Example 7.8. Recall the diagram of the four fundamental subspaces. The indicated direct
sum decompositions of the domain IR
n
and co-domain IR
m
are given easily as follows.
Let X E IR
n
X = PN(A)u + PN(A)X
= A+ Ax + (I - A+ A)x
= VI vt x + V
2
Vi x (recall VVT = I).
Similarly, let y e ]R
m
Example 7.9. Let
Then
and we can decompose the vector [2 3 4]
r
uniquely into the sum of a vector in A/' CA)-
1
and a vector in J\f(A), respectively, as follows:
Definition 7.10. Let V be a vector space over R. Then { , ) : V x V
product if
is a real inner
1. (x, x) > Qfor all x 6V and ( x , x } =0 if and only ifx = 0.
2. (x, y) = (y,x)forallx,y e V.
3. { *, cryi + ^2) = a(x, y\) + / 3( j t, y^} for all jc, yi, j2 ^ V and for alia, ft e R.
Example 7.11. Let V = R". Then { ^, y} = X
T
y is the "usual" Euclidean inner product or
dot product.
Example 7.12. Let V = E". Then ( j c, y)
Q
= X
T
Qy, where Q = Q
T
> 0 is an arbitrary
n x n positive definite matrix, defines a "weighted" inner product.
Definition 7.13. If A e R
mx
", then A
T
e R
nxm
is the unique linear transformation or map
such that (x, Ay) - (A
T
x, y) for all x R
m
and for all y e R".
Similarly, let Y E IR
m
Y = PR(A)Y +
= AA+y + (l- AA+)y
= U1Ur y + U2U[ Y (recall UU
T
= I).
Example 7.9. Let
Then
1/4
1/4
o
1/4 ]
1/4
o
and we can decompose the vector [2 3 4V uniquely into the sum of a vector in N(A)-L
and a vector in N(A), respectively, as follows:
[ ! ] A' Ax + (l - A' A)x
[
1/2 1/2 0] [ 2] [
= ! +
[
5/2] [-1/2]
= + .
1/2
-1/2
o
-1/2
1/2
o
Definition 7.10. Let V be a vector space over IR. Then (', .) : V x V -+ IR is a real inner
product if
1. (x, x) ::: Of or aU x E V and (x, x) = 0 if and only ifx = O.
2. (x, y) = (y, x) for all x, y E V.
3. (x, aYI + PY2) = a(x, Yl) + f3(x, Y2) for all x, Yl, Y2 E V and/or all a, f3 E IR.
Example 7.11. Let V = IRn. Then (x, y) = x
T
Y is the "usual" Euclidean inner product or
dot product.
Example 7.12. Let V = IRn. Then (x, y) Q = X T Qy, where Q = Q T > 0 is an arbitrary
n x n positive definite matrix, defines a "weighted" inner product.
Definition 7.13. If A E IR
m
xn, then ATE IR
n
xm is the unique linear transformation or map
such that {x, Ay) = {AT x, y) for all x E IR
m
andfor all y e IRn.
7.2. Inner Product Spaces 55
It is easy to check that, with this more "abstract" definition of transpose, and if the
(/, y)th element of A is a
(;
, then the (i, y)t h element of A
T
is a/ , . It can also be checked
that all the usual properties of the transpose hold, such as (Afl) = B
T
A
T
. However, the
definition above allows us to extend the concept of transpose to the case of weighted inner
products in the following way. Suppose A e R
mxn
and let {-, -}g and (, -}
R
, with Q and
R positive definite, be weighted inner products on R
m
and W, respectively. Then we can
define the "weighted transpose" A
#
as the unique map that satisfies
(x, Ay)
Q
= (A
#
x, y)
R
for all x e R
m
and for all y e W
1
.
By Example 7.12 above, we must then have X
T
QAy = x
T
(A
#
) Ry for all x, y. Hence we
must have QA = (A
#
) R. Taking transposes (of the usual variety) gives A
T
Q = RA
#
.
Since R is nonsingular, we find
A* = /r'A' Q.
We can also generalize the notion of orthogonality (x
T
y = 0) to Q -orthogonality (Q is
a positive definite matrix). Two vectors x, y e W are <2-orthogonal (or conjugate with
respect to Q) if ( x, y}
Q
= X
T
Qy = 0. Q -orthogonality is an important tool used in
studying conjugate direction methods in optimization theory.
Definition 7.14. Let V be a vector space over C. Then {-, } : V x V -> C is a complex
inner product if
1. ( x, x) > Qfor all x e V and ( x, x) =0 if and only ifx = 0.
2. (x, y) = (y, x) for all x, y e V.
3. (x,ayi + fiy
2
) = a(x, y\) + fi(x, y
2
}forallx, y\, y
2
e V and for alia, ft 6 C.
Remark 7.15. We could use the notation {, -}
c
to denote a complex inner product, but
if the vectors involved are complex-valued, the complex inner product is to be understood.
Note, too, from part 2 of the definition, that ( x, x) must be real for all x.
Remark 7.16. Note from parts 2 and 3 of Definition 7.14 that we have
(ax\ + fix
2
, y) = a(x\, y) + P(x
2
, y}.
Remark 7.17. The Euclidean inner product of x, y e C" is given by
The conventional definition of the complex Euclidean inner product is (x, y} = y
H
x but we
use its complex conjugate x
H
y here for symmetry with the real case.
Remark 7.18. A weighted inner product can be defined as in the real case by (x, y}
Q

X
H
Qy, for arbitrary Q = Q
H
> 0. The notion of Q -orthogonality can be similarly
generalized to the complex case.
7.2. Inner product Spaces 55
It is easy to check that, with this more "abstract" definition of transpose, and if the
(i, j)th element of A is aij, then the (i, j)th element of AT is ap. It can also be checked
that all the usual properties of the transpose hold, such as (AB) = BT AT. However, the
definition above allows us to extend the concept of transpose to the case of weighted inner
products in the following way. Suppose A E ]Rm xn and let (., .) Q and (., .) R, with Q and
R positive definite, be weighted inner products on IR
m
and IRn, respectively. Then we can
define the "weighted transpose" A # as the unique map that satisfies
(x, AY)Q = (A#x, Y)R for all x E IRm and for all Y E IRn.
By Example 7.l2 above, we must then have x
T
QAy = x
T
(A#{ Ry for all x, y. Hence we
must have QA = (A#{ R. Taking transposes (of the usual variety) gives AT Q = RA#.
Since R is nonsingular, we find
A# = R-1A
T
Q.
We can also generalize the notion of orthogonality (x
T
y = 0) to Q-orthogonality (Q is
a positive definite matrix). Two vectors x, y E IRn are Q-orthogonal (or conjugate with
respect to Q) if (x, y) Q = X T Qy = O. Q-orthogonality is an important tool used in
studying conjugate direction methods in optimization theory.
Definition 7.14. Let V be a vector space over <C. Then (., .) : V x V -+ C is a complex
inner product if
1. (x, x) :::: 0 for all x E V and (x, x) = 0 if and only if x = O.
2. (x, y) = (y, x) for all x, y E V.
3. (x, aYI + f3Y2) = a(x, yll + f3(x, Y2) for all x, YI, Y2 E V andfor all a, f3 E c.
Remark 7.15. We could use the notation (., )e to denote a complex inner product, but
if the vectors involved are complex-valued, the complex inner product is to be understood.
Note, too, from part 2 of the definition, that (x, x) must be real for all x.
Remark 7.16. Note from parts 2 and 3 of Definition 7.14 that we have
Remark 7.17. The Euclidean inner product of x, y E C
n
is given by
n
(x, y) = LXiYi = xHy.
i=1
The conventional definition of the complex Euclidean inner product is (x, y) = yH x but we
use its complex conjugate x
H
y here for symmetry with the real case.
Remark 7.1S. A weighted inner product can be defined as in the real case by (x, y)Q =
x
H
Qy, for arbitrary Q = QH > o. The notion of Q-orthogonality can be similarly
generalized to the complex case.
Definition 7.19. A vector space (V, F) endowed with a specific inner product is called an
inner product space. If F = C, we call V a complex inner product space. If F = R, we
call V a real inner product space.
Example 7.20.
1. Check that V = R"
x
" with the inner product (A, B) = Tr A
T
B is a real inner product
space. Note that other choices are possible since by properties of the trace function,
TrA
T
B = TrB
T
A = TrAB
T
= TrBA
T
.
2. Check that V = C
nx
H
B is a complex inner
product space. Again, other choices are possible.
Definition 7.21. Let V be an inner product space. For v e V, we define the norm (or
length) ofv by \\v\\ = */(v, v). This is called the norm induced by ( - , - ) .
Example 7.22.
1. If V = E." with the usual inner product, the induced norm is given by | | i> | | =
xV* 9\ 7
( E , =i < Y )
2
-
2. If V = C" with the usual inner product, the induced norm is given by \\v\\ =
( ?
=
, l , - l
2
)* .
Theorem 7.23. Let P be an orthogonal projection on an inner product space V. Then
\\Pv\\ < \\v\\forallv e V.
Proof: Since P is an orthogonal projection, P
2
= P = P
#
. (Here, the notation P
#
denotes
the unique linear transformation that satisfies ( P u , v } = (u, P
#
v) for all u, v e V. If this
seems a little too abstract, consider V = R" (or C"), where P
#
is simply the usual P
T
(or
P
H
)). Hence ( P v , v) = (P
2
v, v) = (Pv, P
#
v) = ( P v , Pv) = \\Pv\\
2
> 0. Now / - P is
also a projection, so the above result applies and we get
from which the theorem follows.
Definition 7.24. The norm induced on an inner product space by the "usual" inner product
is called the natural norm.
In case V = C" or V = R", the natural norm is also called the Euclidean norm. In
the next section, other norms on these vector spaces are defined. A converse to the above
procedure is also available. That is, given a norm defined by \\x\\ > /( * > x), an inner
product can be defined via the following.
Definition 7.19. A vector space (V, IF) endowed with a specific inner product is called an
inner product space. If IF = e, we call V a complex inner product space. If IF = R we
call V a real inner product space.
Example 7.20.
1. Check that V = IR
n
xn with the inner product (A, B) = Tr AT B is a real inner product
space. Note that other choices are possible since by properties of the trace function,
Tr AT B = Tr B T A = Tr A B T = Tr BAT.
2. Check that V = e
nxn
with the inner product (A, B) = Tr AH B is a complex inner
product space. Again, other choices are possible.
Definition 7.21. Let V be an inner product space. For v E V, we define the norm (or
length) ofv by IIvll = -J(V,V). This is called the norm induced by (', .).
Example 7.22.
1. If V = IR
n
with the usual inner product, the induced norm is given by II v II
n 2 1
(Li=l V
i
)2.
2. If V = en with the usual inner product, the induced norm is given by II v II =
"n 2 !
(L...i=l IVi I ) .
Theorem 7.23. Let P be an orthogonal projection on an inner product space V. Then
IIPvll ::::: Ilvll for all v E V.
Proof: Since P is an orthogonal projection, p2 = P = pH. (Here, the notation p# denotes
the unique linear transformation that satisfies (Pu, v) = (u, p#v) for all u, v E V. If this
seems a little too abstract, consider V = IR
n
(or en), where p# is simply the usual pT (or
pH)). Hence (Pv, v) = (P
2
v, v) = (Pv, p#v) = (Pv, Pv) = IIPvll
2
::: O. Now /- P is
also a projection, so the above result applies and we get
0::::: ((I - P)v. v) = (v. v) - (Pv, v)
= IIvll2 - IIPvll
2
from which the theorem follows. 0
Definition 7.24. The norm induced on an inner product space by the "usual" inner product
is called the natural norm.
In case V = en or V = IR
n
, the natural norm is also called the Euclidean norm. In
the next section, other norms on these vector spaces are defined. A converse to the above
procedure is also available. That is, given a norm defined by IIx II = .j(X,X}, an inner
product can be defined via the following.
7.3. Vector Norms 57
Theorem 7.25 (Polarization Identity).
1. For x, y R", an inner product is defined by
7.3 Vector Norms
Definition 7.26. Let (V, F) be a vector space. Then \ \ - \ \ : V -> R is a vector norm if it
satisfies the following three properties:
2. For x, y e C", an inner product is defined by
where j = i = \/T.
(This is called the triangle inequality, as seen readily from the usual diagram illus
trating the sum of two vectors in R
2
.)
Remark 7.27. It is convenient in the remainder of this section to state results for complex-
valued vectors. The specialization to the real case is obvious.
Definition 7.28. A vector space (V, F) is said to be a normed linear space if and only if
there exists a vector norm || || : V -> R satisfying the three conditions of Definition 7.26.
Example 7.29.
1. For x e C", the Holder norms, or p-norms, are defined by
Special cases:
(The second equality is a theorem that requires proof.)
7.3. Vector Norms
Theorem 7.25 (Polarization Identity).
1. For x, y E an inner product is defined by
(x,y)=xTy=
2. For x, y E en, an inner product is defined by
where j = i = .J=I.
7.3 Vector Norms
IIx + yll2 _ IIxll2 _ lIyll2
2
57
Definition 7.26. Let (V, IF) be a vector space. Then II . II : V ---+ IR is a vector norm ifit
satisfies the following three properties:
1. Ilxll::: Of or all x E V and IIxll = 0 ifand only ifx = O.
2. Ilaxll = lalllxllforallx E Vandforalla E IF.
3. IIx + yll :::: IIxll + IIYliforall x, y E V.
(This is called the triangle inequality, as seen readily from the usual diagram illus-
trating the sum of two vectors in ]R2 .)
Remark 7.27. It is convenient in the remainder of this section to state results for complex-
valued vectors. The specialization to the real case is obvious.
Definition 7.28. A vector space (V, IF) is said to be a normed linear space if and only if
there exists a vector norm II . II : V ---+ ]R satisfying the three conditions of Definition 7.26.
Example 7.29.
1. For x E en, the HOlder norms, or p-norms, are defined by
Special cases:
(a) Ilx III = L:7=1 IXi I (the "Manhattan" norm).
1 1
(b) Ilxllz = (L:7=1Ix;l2)2 = (X
H
X)2 (the Euclidean norm).
(c) Ilxlioo = maxlx;l = lim IIxllp-
IE!! p---++oo
(The second equality is a theorem that requires proof.)
2. Some weighted p-norms:
(a) | | JC| | , .
D
= E^rf/l*/!, where 4 > 0.
(b) I k llz . g (x
h
Q
X
Y > where Q = Q
H
> 0 (this norm is more commonly
denoted || ||
c
).
3. On the vector space (C[to, t \ ] , R), define the vector norm
On the vector space ((C[to, t\])
n
, R), define the vector norm
Fhcorem 7.30 (Holder Inequality). Let x, y e C". Ther,
A particular case of the Holder inequality is of special interest.
Theorem 7.31 (Cauchy-Bunyakovsky-Schwarz Inequality). Let x, y e C". Then
with equality if and only if x and y are linearly dependent.
Proof: Consider the matrix [x y] e C"
x2
. Since
is a nonnegative definite matrix, its determinant must be nonnegative. In other words,
0 < ( x
H
x ) ( y
H
y ) ( x
H
y ) ( y
H
x ) . Since y
H
x = x
H
y, we see immediately that \X
H
y\ <
\\X\\2\\y\\2-
D
Note: This is not the classical algebraic proof of the Cauchy-Bunyakovsky-Schwarz
(C-B-S) inequality (see, e.g., [20, p. 217]). However, it is particularly easy to remember.
Remark 7.32. The angle 0 between two nonzero vectors x, y e C" may be defined by
cos# = I, |.^|| , 0 < 0 < 5-. The C-B-S inequality is thus equivalent to the statement
Il-Mmlylb ^
|COS 0| <1.
Remark 7.33. Theorem 7.31 and Remark 7.32 are true for general inner product spaces.
Remark 7.34. The norm || ||
2
is unitarily invariant, i.e., if U C"
x
" is unitary, then
\\Ux\\
2
= \\x\\
2
(Proof. \\Ux\\l = x
H
U
H
Ux = X
H
X = \\x\\\). However, || - ||, and || - 1^
2. Some weighted p-norms:
(a) IIxll1.D = whered; > O.
1
(b) IIx IIz.Q = (x
H
Qx) 2, where Q = QH > 0 (this norm is more commonly
denoted II . IIQ)'
3. On the vector space (C[to, ttl, 1Ft), define the vector norm
11111 = max 1/(t)I

On the vector space e[to, ttlr, 1Ft), define the vector norm
1111100 = max II/(t) 11
00
,

Theorem 7.30 (HOlder Inequality). Let x, y E en. Then
I I
-+-=1.
p q
A particular case of the HOlder inequality is of special interest.
Theorem 7.31 (Cauchy-Bunyakovsky-Schwarz Inequality). Let x, y E en. Then
with equality if and only if x and yare linearly dependent.
Proof' Consider the matrix [x y] E en
x2
. Since
is a nonnegative definite matrix, its determinant must be nonnegative. In other words,
o (x
H
x)(yH y) - (x
H
y)(yH x). Since yH x = x
H
y, we see immediately that IXH yl
IIxll2l1yllz. 0
Note: This is not the classical algebraic proof of the Cauchy-Bunyakovsky-Schwarz
(C-B-S) inequality (see, e.g., [20, p. 217]). However, it is particularly easy to remember.
Remark 7.32. The angle e between two nonzero vectors x, y E en may be defined by
cos e = 0 e I' The C-B-S inequality is thus equivalent to the statement
1 cose 1 1.
Remark 7.33. Theorem 7.31 and Remark 7.32 are true for general inner product spaces.
Remark 7.34. The norm II . 112 is unitarily invariant, i.e., if U E e
nxn
is unitary, then
IIUxll2 = IIxll2 (Proof IIUxili = XHUHUx = xHx = IIxlli) However, 11111 and 1IIIClO
7.4. Matrix Norms 59
are not unitarily invariant. Similar remarks apply to the unitary invariance of norms of real
vectors under orthogonal transformation.
Remark 7.35. If x, y C" are orthogonal, then we have the Pythagorean Identity
7.4 Matrix Norms
In this section we introduce the concept of matrix norm. As with vectors, the motivation for
using matrix norms is to have a notion of either the size of or the nearness of matrices. The
former notion is useful for perturbation analysis, while the latter is needed to make sense of
"convergence" of matrices. Attention is confined to the vector space (W
nxn
, R) since that is
what arises in the majority of applications. Extension to the complex case is straightforward
and essentially obvious.
Definition 7.39. || || : R
mx
" -> E is a matrix norm if it satisfies the following three
properties:
2 _ _/ / .
the proof of which follows easily from ||z||2 = z z.
Theorem 7.36. All norms on C" are equivalent; i.e., there exist constants c\, c-i (possibly
depending onn) such that
Example 7.37. For x G C", the following inequalities are all tight bounds; i.e., there exist
vectors x for which equality holds:
Finally, we conclude this section with a theorem about convergence of vectors. Con-
vergence of a sequence of vectors to some limit vector can be converted into a statement
about convergence of real numbers, i.e., convergence in terms of vector norms.
Theorem 7.38. Let \\ \\ be a vector norm and suppose v, i
( 1 )
, v
(2
\ ... e C". Then
are not unitarily invariant. Similar remarks apply to the unitary invariance of norms of real
vectors under orthogonal transformation.
Remark 7.35. If x, y E en are orthogonal, then we have the Pythagorean Identity
Ilx = +
the proof of which follows easily from liz = ZH z.
Theorem 7.36. All norms on en are equivalent; i.e., there exist constants CI, C2 (possibly
depending on n) such that
Example 7.37. For x E en, the following inequalities are all tight bounds; i.e., there exist
vectors x for which equality holds:
Ilxlll :::: Jn Ilxlb
Ilxll2:::: IIxll
IIxlloo :::: IIxll
Ilxlll :::: n IIxlloo;
IIxl12 :::: Jn Ilxll
oo
;
IIxlioo :::: IIxllz.
Finally, we conclude this section with a theorem about convergence of vectors. Con-
vergence of a sequence of vectors to some limit vector can be converted into a statement
about convergence of real numbers, i.e., convergence in terms of vector norms.
Theorem 7.38. Let II II be a vector norm and suppose v, v(l), v(2), ... E en. Then
lim V(k) = v if and only if lim II v(k) - v II = O.
k4+00
7.4 Matrix Norms
In this section we introduce the concept of matrix norm. As with vectors, the motivation for
using matrix norms is to have a notion of either the size of or the nearness of matrices. The
former notion is useful for perturbation analysis, while the latter is needed to make sense of
"convergence" of matrices. Attention is confined to the vector space (IRm xn , IR) since that is
what arises in the majority of applications. Extension to the complex case is straightforward
and essentially obvious.
Definition 7.39. II II : IR
mxn
IR is a matrix norm if it satisfies the following three
properties:
1. IIAII Of or all A E IR
mxn
and IIAII = 0 if and only if A = O.
2. lIaAl1 = lalliAliforall A E IR
mxn
andfor all a E IR.
3. IIA + BII :::: IIAII + IIBII for all A, BE IRmxn.
(As with vectors, this is called the triangle inequality.)
mx
". Then the Frobenius norm (or matrix Euclidean norm) is
defined by
^wncic r = laiiK^/i;;.
mxn
. Then the matrix p-norms are defined by
The following three special cases are important because they are "computable." Each is a
theorem and requires a proof.
1. The "maximum column sum" norm is
2. The "maximum row sum" norm is
3. The spectral norm is
Example 7.42. Let A E R
mxn
. The Schatten/7-norms are defined by
Some special cases of Schatten /?-norms are equal to norms defined previously. For example,
|| . ||
5 2
= || . \\
F
and || ||
5i00
= || ||
2
. The norm || ||
5>1
is often called the trace norm.
Example 7.43. Let A e K
mx
". Then "mixed" norms can also be defined by
Example 7.44. The "matrix analogue of the vector 1-norm," || A\\
s
= ^ j \a
i}
; |, is a norm.
The concept of a matrix norm alone is not altogether useful since it does not allow us
to estimate the size of a matrix product A B in terms of the sizes of A and B individually.
Example 7.40. Let A E lR,mxn. Then the Frobenius norm (or matrix Euclidean norm) is
defined by
IIAIIF ~ (t. t ai;) I ~ (t. altA)) 1 (T, (A' A)) 1 (T, (AA '));
(where r = rank(A)).
Example 7.41. Let A E lR,mxn. Then the matrix p-norms are defined by
IIAxll
IIAII = max -_P = max IIAxll .
P Ilxllp;60 Ilxli
p
IIxllp=1 p
The following three special cases are important because they are "computable." Each is a
theorem and requires a proof.
I. The "maximum column sum" norm is
2. The "maximum row sum" norm is
IIAlioo = max
rE!!l. (
t laUI ).
J=1
3. The spectral norm is
tTL T
IIAII2 = Amax(A A) = A ~ a x ( A A ) = a1(A).
Note: IIA+llz = l/ar(A), where r = rank(A).
Example 7.42. Let A E lR,mxn. The Schattenp-norms are defined by
I
IIAlls.p = (at' + ... + a!)"".
Some special cases of Schatten p-norms are equal to norms defined previously. For example,
11115.2 = II . IIF and 11'115,00 = II . 112' The norm II . 115.1 is often called the trace norm.
Example 7.43. Let A E lR,mxn _ Then "mixed" norms can also be defined by
IIAII = max IIAxil
p
p,q 11.<110#0 IIxllq
Example 7.44. The "matrix analogue of the vector I-norm," IIAlis = Li.j laij I, is a norm.
The concept of a matrix norm alone is not altogether useful since it does not allow us
to estimate the size of a matrix product AB in terms of the sizes of A and B individually.
Notice that this difficulty did not arise for vectors, although there are analogues for, e.g.,
inner products or outer products of vectors. We thus need the following definition.
Definition 7.45. Let A e R
mxn
, B e R
nxk
. Then the norms \\ \\
a
, \\ \\
p
, and \\ \\
y
are
mutually consistent if \\ A B \\
a
< \\A\\p\\B\\
y
. A matrix norm\\ \\ is said to be consistent
if \\AB\\ < || A || || fi|| whenever the matrix product is defined.
Example 7.46.
1. || ||/7 and || ||
p
for all p are consistent matrix norms.
2. The "mixed" norm
is a matrix norm but it is not consistent. For example, take A = B = \ \ J1. Then
| | Af l | |
l i 00
= 2whil e| | A| |
l i 00
| | B| |
1 >00
= l.
The p -norms are examples of matrix norms that are subordinate to (or induced by)
a vector norm, i.e.,
11^ 4^ 11
(or, more generally, ||A|| = maxô ., . .
p
) . For such subordinate norms, also called oper-
ator norms, we clearly have ||Aj c|| < ||A||1|jt||. Since | | Af ij c| | < | | A| | | | f l j c| | < ||A||||fl||||j t||,
it follows that all subordinate norms are consistent.
Theorem 7.47. There exists a vector x* such that ||Ajt*|| = ||A|| ||jc*|| if the matrix normis
subordinate to the vector norm.
Theorem 7.48. If \\ \\
m
is a consistent matrix norm, there exists a vector norm \\ \\
v
consistent with it, i.e., H Aj c JI ^ < \\A\\
m
\\x\\
v
.
Not every consistent matrix norm is subordinate to a vector norm. For example,
consider || \\
F
. Then| | A^| |
2
< ||A||
F
||j c||
2
, so || ||
2
is consistent with || ||
F
, but there does
not exist a vector norm || || such that ||A||
F
is given by maxô \^
Useful Results
The following miscellaneous results about matrix norms are collected for future reference.
The interested reader is invited to prove each of them as an exercise.
2. For A e R"
x
", the following inequalities are all tight, i.e., there exist matrices A for
which equality holds:
Notice that this difficulty did not arise for vectors, although there are analogues for, e.g.,
inner products or outer products of vectors. We thus need the following definition.
Definition 7.45. Let A E ]Rmxn, B E ]Rnxk. Then the norms II . II", II Ilfl' and II . lIy are
mutuallyconsistentifIlABII,,::S IIAllfllIBlly. A matrix norm 1111 is said to be consistent
if II A B II ::s II A 1111 B II whenever the matrix product is defined.
Example 7.46.
1. II II F and II . II p for all p are consistent matrix norms.
2. The "mixed" norm
IIAxll1
II 11
100
= max --= max laijl
, x;60 Ilx 1100 i,j
is a matrix norm but it is not consistent. For example, take A = B = [: :]. Then
IIABIII,oo = 2 while IIAIII,ooIlBIII,oo = 1.
The p-norms are examples of matrix norms that are subordinate to (or induced by)
a vector norm, i.e.,
IIAxl1
IIAII = max -- = max IIAxl1
x;60 IIx II Ilxll=1
IIAxll .
(or, more generally, IIAllp,q = maxx;60 IIxll
q
P
), For such subordmate norms, also caUedoper-
atornorms, wec1earlyhave IIAxll ::s IIAllllxll Since IIABxl1 ::s IIAlIllBxll ::s IIAIIIIBllllxll,
it follows that all subordinate norms are consistent.
Theorem 7.47. There exists a vector x* such that IIAx*11 = IIAllllx*11 if the matrix norm is
subordinate to the vector norm.
Theorem 7.48. If II . 11m is a consistent matrix norm, there exists a vector norm II . IIv
consistent with it, i.e., IIAxliv ::s IIAlim Ilxli
v
'
Not every consistent matrix norm is subordinate to a vector norm. For example,
consider II . II F' Then II Ax 112 ::s II A II Filx 112, so II . 112 is consistent with II . II F, but there does
not exist a vector norm II . II such that IIAIIF is given by max
x
;60 " , ~ ~ i ' .
Useful Results
The following miscellaneous results about matrix norms are collected for future reference.
The interested reader is invited to prove each of them as an exercise.
1. II In II p = 1 for all p, while IIIn II F = .jii.
2. For A E ]Rnxn, the following inequalities are all tight, i.e., there exist matrices A for
which equality holds:
IIAIII ::s .jii IIAlb
IIAII2 ::s.jii IIAII
I
,
II A 1100 ::s n IIAII
I
,
IIAIIF ::s.jii IIAII
I
,
IIAIII ::s n IIAlloo,
IIAII2 ::s .jii IIAlloo,
IIAlioo ::s .jii IIAII2,
IIAIIF ::s .jii IIAlb
IIAIII ::s .jii II
A
IIF;
IIAII2::S IIAIIF;
IIAlioo ::s .jii IIAIIF;
IIAIIF ::s .jii IIAlioo'
3. For A eR
mxa
.
4. The norms || \\
F
and || ||
2
(as well as all the Schatten /?-norms, but not necessarily
other p-norms) are unitarily invariant; i.e., for all A e R
mx
" and for all orthogonal
matrices Q zR
mxm
and Z e M"
x
", ||(MZ||
a
= | | A| |
a
fora = 2 or F.
Convergence
The following theorem uses matrix norms to convert a statement about convergence of a
sequence of matrices into a statement about the convergence of an associated sequence of
scalars.
Theorem 7.49. Let \\ -\\bea matrix normand suppose A, A
( 1)
, A
(2)
, ... e R
mx
". Then
EXERCISES
1. If P is an orthogonal projection, prove that P
+
= P.
2. Suppose P and Q are orthogonal projections and P + Q = I. Prove that P Q
must be an orthogonal matrix.
3. Prove that / A
+
A is an orthogonal projection. Also, prove directly that V
2
V/ is an
orthogonal projection, where 2 is defined as in Theorem 5.1.
4. Suppose that a matrix A e W
nxn
has linearly independent columns. Prove that the
orthogonal projection onto the space spanned by these column vectors is given by the
matrix P = A(A
T
A)~
}
A
T
.
5. Find the (orthogonal) projection of the vector [2 3 4]
r
onto the subspace of R
3
spanned by the plane 3;c v + 2z = 0.
6. Prove that E"
x
T
B is a real inner product
space.
7. Show that the matrix norms || ||
2
and || \\
F
are unitarily invariant.
8. Definition: Let A e R
nxn
and denote its set of eigenvalues (not necessarily distinct)
by { A-i , . . . , > . } . The spectral radius of A is the scalar
3. For A E IR
mxn
,
max laijl :::: IIAII2 :::: ~ max laijl.
l.] l.]
4. The norms II . IIF and II . 112 (as well as all the Schatten p-norms, but not necessarily
other p-norms) are unitarily invariant; i.e., for all A E IR
mxn
and for all orthogonal
matrices Q E IR
mxm
and Z E IR
nxn
, IIQAZlia = IIAlla fora = 2 or F.
Convergence
The following theorem uses matrix norms to convert a statement about convergence of a
sequence of matrices into a statement about the convergence of an associated sequence of
scalars.
Theorem 7.49. Let II 11 be a matrix norm and suppose A, A(I), A(2), ... E IRmxn. Then
lim A (k) = A if and only if lim IIA (k) - A II = o.
k ~ + o o k ~ + o o
EXERCISES
1. If P is an orthogonal projection, prove that p+ = P.
2. Suppose P and Q are orthogonal projections and P + Q = I. Prove that P - Q
must be an orthogonal matrix.
3. Prove that I - A + A is an orthogonal projection. Also, prove directly that V
2
Vl is an
orthogonal projection, where V2 is defined as in Theorem 5.1.
4. Suppose that a matrix A E IR
mxn
has linearly independent columns. Prove that the
orthogonal projection onto the space spanned by these column vectors is given by the
matrix P = A(AT A) -1 AT.
5. Find the (orthogonal) projection of the vector [2 3 4f onto the subspace of 1R
3
spanned by the plane 3x - y + 2z = O.
6. Prove that IR
n
xn with the inner product (A, B) = Tr AT B is a real inner product
space.
7. Show that the matrix norms II . 112 and II . IIF are unitarily invariant.
8. Definition: Let A E IR
nxn
and denote its set of eigenvalues (not necessarily distinct)
by P.l, ... , An}. The spectral radius of A is the scalar
p(A) = max IA;I.
i
Exercises 63
Determine ||A||
F
, H AI d , ||A||
2
, H AH ^ , and p(A). (An n x n matrix, all of whose
columns and rows as well as main d iagonal and antid iagonal sum to s = n(n
2
+ l)/2,
is called a "magic square" matrix. I f M is a magic square matrix, it can be proved
that || M U p = s for all/?.)
10. Let A = xy
T
, where both x, y e R" are nonzero. Determine ||A||
F
, ||A||j, ||A||
2
,
and ||A||oo in terms of \\x\\
a
and /or \\y\\p, where a and ft take the value 1, 2, or oo as
appropriate.
Let
9. Let
Determine ||A||
F
, \\A\\
lt
||A||
2
, H A^ , and p(A).
Exercises 63
Let
A = [ ~ 0 ~ ] .
14 12 5
Determine IIAIIF' IIAII
I
, IIAlb IIAlloo, and peA).
9. Let
A = [ ~ ~ ~ ] .
492
Determine IIAIIF' IIAII
I
, IIAlb IIAlloo, and peA). (An n x n matrix, all of whose
columns and rows as well as main diagonal and antidiagonal sum to s = n (n
2
+ 1) /2,
is called a "magic square" matrix. If M is a magic square matrix, it can be proved
that IIMllp = s for all p.)
10. Let A = xyT, where both x, y E IR
n
are nonzero. Determine IIAIIF' IIAIII> IIAlb
and II A 1100 in terms of IIxlla and/or IlylljJ, where ex and {3 take the value 1,2, or (Xl as
appropriate.
Chapter 8
Li near Least Squares
Problems
8.1 The Li near Least Squares Problem
Problem: Suppose A e R
mx
" with m > n and b <= R
m
is a given vector. The linear least
squares problem consists of finding an element of the set
Solution: The set X has a number of easily verified properties:
1. A vector x e X if and only if A
T
r = 0, where r = b Ax is the residual associated
with x. The equations A
T
r 0 can be rewritten in the form A
T
Ax = A
T
b and the
latter form is commonly known as the normal equations, i.e., x e X if and only if
x is a solution of the normal equations. For further details, see Section 8.2.
2. A vector x E X if and onlv if x is of the form
To see why this must be so, write the residual r in the form
Now, (Pn(A)b AJ C ) is clearly in 7(A) , while
so these two vectors are orthogonal. Hence,
from the Pythagorean identity (Remark 7.35). Thus, ||A.x b\\\ (and hence p ( x ) =
\\Ax b\\2) assumes its minimum value if and only if
65
Chapter 8
Linear Least Squares
Problems
8.1 The Linear Least Squares Problem
Problem: Suppose A E jRmxn with m 2: nand b E jRm is a given vector. The linear least
squares problem consists of finding an element of the set
x = {x E jRn : p(x) = IIAx - bll
2
is minimized}.
Solution: The set X has a number of easily verified properties:
1. A vector x E X if and only if AT r = 0, where r = b - Ax is the residual associated
with x. The equations AT r = 0 can be rewritten in the form A T Ax = AT b and the
latter form is commonly known as the normal equations, i.e., x E X if and only if
x is a solution of the normal equations. For further details, see Section 8.2.
2. A vector x E X if and only if x is of the form
x=A+b+(I-A+A)y, whereyEjRnisarbitrary. (8.1)
To see why this must be so, write the residual r in the form
r = (b - PR(A)b) + (PR(A)b - Ax).
Now, (PR(A)b - Ax) is clearly in 'R(A), while
(b - PR(A)b) = (I - PR(A))b
= PR(A),,-b E 'R(A)-L
so these two vectors are orthogonal. Hence,
= lib -
= lib - + IIPR(A)b -
from the Pythagorean identity (Remark 7.35). Thus, IIAx - (and hence p(x) =
II Ax - b 112) assumes its minimum value if and only if
(8.2)
65
66 Chapter 8. Linear Least Squares Problems
and this equation always has a solution since AA
+
b e 7(A). By Theorem 6.3, all
solutions of (8.2) are of the form
where y e W is arbitrary. The minimum value of p ( x ) is then clearly equal to
the last inequality following by Theorem 7.23.
3. X is convex. To see why, consider two arbitrary vectors jci = A
+
b + (I A+A)y
and *2 = A+b + (I A+A)z in X. Let 6 e [0, 1]. Then the convex combination
0*i + (1 - #)*
2
= A+b + (I - A
+
A)(Oy + (1 - 0)z) is clearly in X.
4. X has a unique element x* of minimal 2-norm. In fact, x* = A
+
b is the unique vector
that solves this "double minimization" problem, i.e., x * minimizes the residual p ( x )
and is the vector of minimum 2-norm that does so. This follows immediately from
convexity or directly from the fact that all x e X are of the form (8.1) and
which follows since the two vectors are orthogonal.
5. There is a unique solution to the least squares problem, i.e., X = {x*} = {A+b}, if
and only if A
+
A = I or, equivalently, if and only if rank (A) = n.
Just as for the solution of linear equations, we can generalize the linear least squares
problem to the matrix case.
Theorem 8.1. Let A e E
mx
" and B R
mxk
. The general solution to
is of the form
where Y R"
xfc
is arbitrary. The unique solution of minimum 2-norm or F-norm is
X = A+B.
Remark 8.2. Notice that solutions of the linear least squares problem look exactly the
same as solutions of the linear system AX = B. The only difference is that in the case
of linear least squares solutions, there is no "existence condition" such as K(B) c 7(A).
If the existence condition happens to be satisfied, then equality holds and the least squares
and this equation always has a solution since AA+b E R(A). By Theorem 6.3, all
solutions of (8.2) are of the form
x = A+ AA+b + (I - A+ A)y
=A+b+(I-A+A)y,
where y E ]R.n is arbitrary. The minimum value of p (x) is then clearly equal to
lib - PR(A)bll
z
= 11(1 - AA+)bI1
2
~ Ilbll z,
the last inequality following by Theorem 7.23.
3. X is convex. To see why, consider two arbitrary vectors Xl = A + b + (I - A + A) y
and Xz = A+b + (I - A+ A)z in X. Let 8 E [0,1]. Then the convex combination
8x, + (1 - 8)xz = A+b + (I - A+ A)(8y + (1 - 8)z) is clearly in X.
4. X has a unique element x" of minimal2-norm. In fact, x" = A + b is the unique vector
that solves this "double minimization" problem, i.e., x* minimizes the residual p(x)
and is the vector of minimum 2-norm that does so. This follows immediately from
convexity or directly from the fact that all x E X are of the form (8.1) and
which follows since the two vectors are orthogonal.
5. There is a unique solution to the least squares problem, i.e., X = {x"} = {A+b}, if
and only if A + A = lor, equivalently, if and only if rank(A) = n.
Just as for the solution of linear equations, we can generalize the linear least squares
problem to the matrix case.
Theorem 8.1. Let A E ]R.mxn and BE ]R.mxk. The general solution to
min IIAX - Bib
XElR
Plxk
is of the form
X=A+B+(I-A+A)Y,
where Y E ]R.nxk is arbitrary. The unique solution of minimum 2-norm or F-norm is
X = A+B.
Remark 8.2. Notice that solutions of the linear least squares problem look exactly the
same as solutions of the linear system AX = B. The only difference is that in the case
of linear least squares solutions, there is no "existence condition" such as R(B) S; R(A).
If the existence condition happens to be satisfied. then equality holds and the least squares
residual is 0. Of all solutions that give a residual of 0, the unique solution X = A
+
B has
minimum 2-norm or F-norm.
Remark 8.3. If we take B = I
m
in Theorem 8.1, then X = A
+
can be interpreted as
saying that the Moore-Penrose pseudoinverse of A is the best (in the matrix 2-norm sense)
matrix such that AX approximates the identity.
Remark 8.4. Many other interesting and useful approximation results are available for the
matrix 2-norm (and F-norm). One such is the following. Let A e M
x
" with SVD
8.2 Geometric Solution
Looking at the schematic provided in Figure 8.1, it is apparent that minimizing || Ax b\\
2
is equivalent to finding the vector x e W
1
for which p Ax is closest to b (in the Euclidean
norm sense). Clearly, r = b Ax must be orthogonal to 7(A). Thus, if Ay is an arbitrary
vector in 7(A) (i.e., y is arbitrary), we must have
Then a best rank k approximation to A for l <f c <r , i . e . , a solution to
is given by
The special case in which m = n and k = n 1 gives a nearest singular matrix to A e
Since y is arbitrary, we must have A
T
b A
T
Ax = 0 or A
r
A;c = A
T
b.
Special case: If A is full (column) rank, then x = (A
T
A) A
T
b.
8.3 Linear Regression and Other Linear Least Squares
Problems
8.3.1 Example: Linear regression
Suppose we have m measurements (t\,y\), . . . , (t
m
,y
m
) for which we hypothesize a linear
(affine) relationship
residual is O. Of all solutions that give a residual of 0, the unique solution X = A + B has
minimum 2-norm or F -norm.
Remark 8.3. If we take B = 1m in Theorem 8.1, then X = A+ can be interpreted as
saying that the Moore-Penrose pseudoinverse of A is the best (in the matrix 2-norm sense)
matrix such that AX approximates the identity.
Remark 8.4. Many other interesting and useful approximation results are available for the
matrix 2-norm (and F -norm). One such is the following. Let A E with SVD
A = = LOiUiV!.
i=l
Then a best rank k approximation to A for 1 :s k :s r, i.e., a solution to
min IIA - MIi2,
MEJRZ'xn
is given by
k
Mk = LOiUiV!.
i=1
The special case in which m = nand k = n - 1 gives a nearest singular matrix to A E x n .
8.2 Geometric Solution
Looking at the schematic provided in Figure 8.1, it is apparent that minimizing IIAx - bll
2
is equivalent to finding the vector x E lR
n
for which p = Ax is closest to b (in the Euclidean
norm sense). Clearly, r = b - Ax must be orthogonal to R(A). Thus, if Ay is an arbitrary
vector in R(A) (i.e., y is arbitrary), we must have
0= (Ay)T (b - Ax)
=yTAT(b-Ax)
= yT (ATb _ AT Ax).
Since y is arbitrary, we must have AT b - AT Ax = 0 or AT Ax = AT b.
Special case: If A is full (column) rank, then x = (AT A)-l ATb.
8.3 Linear Regression and Other Linear Least Squares
Problems
8.3.1 Example: Linear regression
Suppose we have m measurements (ll, YI), ... , (trn, Ym) for which we hypothesize a linear
(affine) relationship
y = at + f3
(8.3)
Figure 8.1. Projection of b on K(A).
for certain constants a. and ft. One way to solve this problem is to find the line that best fits
the data in the least squares sense; i.e., with the model (8.3), we have
where &\,..., 8
m
are "errors" and we wish to minimize 8\ + + 8^- Geometrically, we
are trying to find the best line that minimizes the (sum of squares of the) distances from the
given data points. See, for example, Figure 8.2.
Figure 8.2. Simple linear regression.
Note that distances are measured in the vertical sense from the points to the line (as
indicated, for example, for the point (t\, y\}}. However, other criteria arc possible. For ex-
ample, one could measure the distances in the horizontal sense, or the perpendicular distance
from the points to the line could be used. The latter is called total least squares. Instead
of 2-norms, one could also use 1-norms or oo-norms. The latter two are computationally
b
r
p=Ax Ay E R(A)
Figure S.l. Projection of b on R(A).
for certain constants a and {3. One way to solve this problem is to find the line that best fits
the data in the least squares sense; i.e., with the model (8.3), we have
YI = all + {3 + 81,
Y2 = al2 + {3 + 82
where 8
1
, ... , 8
m
are "errors" and we wish to minimize 8? + ... + 8;. Geometrically, we
are trying to find the best line that minimizes the (sum of squares of the) distances from the
given data points. See, for example, Figure 8.2.
y
Figure 8.2. Simple linear regression.
Note that distances are measured in the venical sense from the point!; to [he line (a!;
indicated. for example. for the point (tl. YIn. However. other criteria nrc For cx-
ample, one could measure the distances in the horizontal sense, or the perpendiculnr distance
from the points to the line could be used. The latter is called total least squares. Instead
of 2-norms, one could also use I-norms or oo-norms. The latter two are computationally
8.3. Linear Regression and Other Linear Least Squares Problems 69
much more difficult to handle, and thus we present only the more tractable 2-norm case in
text that follows.
The ra "error equations" can be written in matrix form as
where
We then want to solve the problem
or, equivalently,
Solution: x [^1 is a solution of the normal equations A
T
Ax = A
T
y where, for the
special form of the matrices above, we have
and
8.3.2 Other least squares problems
Suppose the hypothesized model is not the linear equation (8.3) but rather is of the form
y = f ( t ) =
Cl
0!(0+ 4- c
n
<t>
n
(t). (8.5)
In (8.5) the < / > ,(0 are given (basis) functions and the c
;
are constants to be determined to
minimize the least squares error. The matrix problem is still (8.4), where we now have
An important special case of (8.5) is least squares polynomial approximation, which
corresponds to choosing 0, (?) = t'~
l
, i
;
e n, although this choice can lead to computational
The solution for the parameters a and ft can then be written
8.3. Linear Regression and Other Linear Least Squares Problems 69
much more difficult to handle, and thus we present only the more tractable 2-norm case in
text that follows.
The m "error equations" can be written in matrix form as
Y = Ax +0,
where
We then want to solve the problem
minoT 0 = min (Ax - y)T (Ax - y)
x
or, equivalently,
min = min II Ax -
x
Solution: x = is a solution of the normal equations AT Ax
special form of the matrices above, we have
and
AT Y = [ Li ti Yi J.
LiYi
The solution for the parameters a and f3 can then be written
8.3.2 Other least squares problems
(8.4)
AT y where, for the
Suppose the hypothesized model is not the linear equation (S.3) but rather is of the form
(8.5)
In (8.5) the i(t) are given (basis) functions and the Ci are constants to be determined to
minimize the least squares error. The matrix problem is still (S.4), where we now have
An important special case of (8.5) is least squares polynomial approximation, which
corresponds to choosing i (t) = t
i
-
1
, i E !!, although this choice can lead to computational
difficulties because of numerical ill conditioning for large n. Numerically better approaches
are based on orthogonal polynomials, piecewise polynomial functions, splines, etc.
The key feature in (8.5) is that the coefficients c, appear linearly. The basis functions
< / > ,- can be arbitrarily nonlinear. Sometimes a problem in which the c, 's appear nonlinearly
can be converted into a linear problem. For example, if the fitting function is of the form
y = f ( t ) = c\e
C2i
, then taking logarithms yields the equation logy = logci + cjt. Then
defining y logy, c\ = logci, and GI = cj_ results in a standard linear least squares
problem.
In the numerical linear algebra literature (e.g., [4], [7], [11], [23]), it is shown that solution
of linear least squares problems via the normal equations can be a very poor numerical
method in finite- precision arithmetic. Since the standard Kalman filter essentially amounts
to sequential updating of normal equations, it can be expected to exhibit such poor numerical
behavior in practice (and it does). Better numerical methods are based on algorithms that
work directly and solely on A itself rather than A
T
A. Two basic classes of algorithms are
based on S VD and QR (orthogonal- upper triangular) factorization, respectively. The former
is much more expensive but is generally more reliable and offers considerable theoretical
insight.
In this section we investigate solution of the linear least squares problem
The last equality follows from the fact that if v = [ ], then ||u||^ = | | i> i \\\ + \\vi\\\ (note
that orthogonality is not what is used here; the subvectors can have different lengths). This
explains why it is convenient to work above with the square of the norm rather than the
norm. As far as the minimization is concerned, the two are equivalent. In fact, the last
quantity above is clearly minimized by taking z\ = S~
l
c\. The subvector z
2
is arbitrary,
while the minimum value of \\Ax b\\^ is l ^l l r
via the SVD. Specifically, we assume that A has an SVD given by A = UT, V
T
= U\SVf
as in Theorem 5.1. We now note that
difficulties because of numerical ill conditioning for large n. Numerically better approaches
are based on orthogonal polynomials, piecewise polynomial functions, splines, etc.
The key feature in (8.5) is that the coefficients Ci appear linearly. The basis functions
i can be arbitrarily nonlinear. Sometimes a problem in which the Ci'S appear nonlinearly
can be converted into a linear problem. For example, if the fitting function is of the form
Y = f (t) = c, e
C2
/ , then taking logarithms yields the equation log y = log c, + c2f. Then
defining y = log y, c, = log c" and C2 = C2 results in a standard linear least squares
problem.
In the numerical linear algebra literature (e.g., [4], [7], [11], [23]), it is shown that solution
of linear least squares problems via the normal equations can be a very poor numerical
method in finite-precision arithmetic. Since the standard Kalman filter essentially amounts
to sequential updating of normal equations, it can be expected to exhibit such poor numerical
behavior in practice (and it does). Better numerical methods are based on algorithms that
work directly and solely on A itself rather than AT A. Two basic classes of algorithms are
based on SVD and QR (orthogonal-upper triangular) factorization, respectively. The former
is much more expensive but is generally more reliable and offers considerable theoretical
insight.
In this section we investigate solution of the linear least squares problem
min II Ax - b11
2
, A E IR
mxn
, bE IR
m
, (8.6)
x
via the SVD. Specifically, we assume that A has an SVD given by A = = U,SVr
as in Theorem 5.1. We now note that
IIAx - = x -
= II VT X - U
T
bll; since II . Ib is unitarily invariant
wherez=VTx,c=UTb
= II [ ] - [ ] II:
= II [ c, ] II:
The last equality follows from the fact that if v = then II v II = II viii + II v211 (note
that orthogonality is not what is used here; the subvectors can have different lengths). This
explains why it is convenient to work above with the square of the norm rather than the
norm. As far as the minimization is concerned. the two are equivalent. In fact. the last
quantity above is clearly minimized by taking z, = S-'c,. The subvector Z2 is arbitrary,
while the minimum value of II Ax - b II is II czll
8.5. Least Squares and QR Factorization 71
Now transform back to the original coordinates:
The last equality follows from
Note that since 12 is arbitrary, V
2
z
2
is an arbitrary vector in 7Z(V
2
) = A/"(A). Thus, x has
been written in the form x = A
+
b + (/ A
+
A ) _ y, where y e R
m
is arbitrary. This agrees,
of course, with (8.1).
The minimum value of the least squares residual is
and we clearly have that
minimum least squares residual is 0 -4=> b is orthogonal to all vectors in U
2
<=^ b is orthogonal to all vectors in 7l(A}
L
Another expression for the minimum residual is || (/ AA
+
) b| |
2
. This follows easily since
||(7 - AA+)b\\
2
2
- \\U2Ufb\\l = b
T
U
2
UÛ
2
UJb = b
T
U
2
U*b = \\U?b\\
2
2
.
Finally, an important special case of the linear least squares problem is the
so-called full-rank problem, i.e., A e 1R
X
" . In this case the SVD of A is given by
A = UZV
T
= [U
{
t/ 2][o]î
r
> and there is thus "no V
2
part" to the solution.
8.5 Least Squares and QR Factorization
In this section, we again look at the solution of the linear least squares problem (8.6) but this
time in terms of the QR factorization. This matrix factorization is much cheaper to compute
than an SVD and, with appropriate numerical enhancements, can be quite reliable.
To simplify the exposition, we add the simplifying assumption that A has full column
rank, i.e., A e R
X M
. It is then possible, via a sequence of so-called Householder or Givens
transformations, to reduce A in the following way. A finite sequence of simple orthogonal
row transformations (of Householder or Givens type) can be performed on A to reduce it
to triangular form. If we label the product of such orthogonal row transformations as the
orthogonal matrix Q
T
R
mxm
, we have
B.S. Least Squares and QR Factorization 71
Now transform back to the original coordinates:
x = Vz
= [VI V
2
1 [ ]
= VIZI + V2Z2
= VIS-ici + V2Z2
= vls-Iufb + V
2
Z
2
.
The last equality follows from
c = U T b = [ f: ] = [ l
Note that since Z2 is arbitrary, V
2
Z
2
is an arbitrary vector in R(V
2
) = N(A). Thus, x has
been written in the form x = A + b + (I - A + A) y, where y E ffi.m is arbitrary. This agrees,
of course, with (8.1).
The minimum value of the least squares residual is
and we clearly have that
minimum least squares residual is 0 {::=:} b is orthogonal to all vectors in U2
{::=:} b is orthogonal to all vectors in R(A)l.
{::=:} b E R(A).
Another expression for the minimum residual is II (I - AA +)bllz. This follows easily since
11(1- = = b
T
U
Z
V!V
2
V!b = bTVZV!b =
Finally, an important special case of the linear least squares problem is the
so-called full-rank problem, i.e., A E In this case the SVD of A is given by
A = V:EV
T
= [VI Vzl[g]Vr, and there is thus "no V
2
part" to the solution.
8.5 Least Squares and QR Factorization
In this section, we again look at the solution of the linear least squares problem (8.6) but this
time in terms of the QR factorization. This matrix factorization is much cheaper to compute
than an SVD and, with appropriate numerical enhancements, can be quite reliable.
To simplify the exposition, we add the simplifying assumption that A has full column
rank, i.e., A E It is then possible, via a sequence of so-called Householder or Givens
transformations, to reduce A in the following way. A finite sequence of simple orthogonal
row transformations (of Householder or Givens type) can be performed on A to reduce it
to triangular form. If we label the product of such orthogonal row transformations as the
orthogonal matrix QT E ffi.mxm, we have
(8.7)
where R e M
x
" is upper triangular. Now write Q = [Q\ Q
2
], where Q\ e R
mx
" and
Q
2
K"
IX(m
~"
)
. Both Q\ and <2
2
have orthonormal columns. Multiplying through by Q
in (8.7), we see that
Any of (8.7), (8.8), or (8.9) are variously referred to as QR factorizations of A. Note that
(8.9) is essentially what is accomplished by the Gram-Schmidt process, i.e., by writing
AR~
l
= Q\ we see that a "triangular" linear combination (given by the coefficients of
R~
l
) of the columns of A yields the orthonormal columns of Q\.
Now note that
The last quantity above is clearly minimized by taking x = R
l
c\ and the minimum residual
is \\C 2\\2- Equivalently, we have x = R~
l
Q\b = A
+
b and the minimum residual is IIC?^!^-
EXERCISES
1. For A W
xn
, b e E
m
, and any y e R", check directly that (I - A
+
A)y and A
+
b
are orthogonal vectors.
2. Consider the following set of measurements (*,, y
t
):
(a) Find the best (in the 2-norm sense) line of the form y = ax + ft that fits this
data.
(b) Find the best (in the 2-norm sense) line of the form jc = ay + (3 that fits this
data.
3. Suppose qi and q
2
are two orthonormal vectors and b is a fixed vector, all in R".
(a) Find the optimal linear combination aq^ + fiq
2
that is closest to b (in the 2-norm
sense).
(b) Let r denote the "error vector" b ctq\ flq
2
- Show that r is orthogonal to
bothî and q
2
.
where R E is upper triangular. Now write Q = [QI Qz], where QI E ffi.mxn and
Qz E ffi.m x (m-n). Both Q I and Qz have orthonormal columns. Multiplying through by Q
in (8.7), we see that

(8.8)
= [QI Qz] [ ]
= QIR.
(8.9)
Any of (8.7), (8.8), or (8.9) are variously referred to as QR factorizations of A. Note that
(8.9) is essentially what is accomplished by the Gram-Schmidt process, i.e., by writing
AR-
1
= QI we see that a "triangular" linear combination (given by the coefficients of
R-
I
) of the columns of A yields the orthonormal columns of Q I.
Now note that
IIAx - = IIQ
T
Ax - since II . 112 is unitarily invariant
= II [ ] x - [ ] If:,
The last quantity above is clearly minimized by taking x = R-
I
Cl and the minimum residual
is Ilczllz. Equivalently, we have x = R-
1
Qf b = A +b and the minimum residual is II Qr bllz'
EXERCISES
1. For A E ffi.
mxn
, b E ffi.
m
, and any y E ffi.
n
, check directly that (I - A + A)y and A +b
are orthogonal vectors.
2. Consider the following set of measurements (Xi, Yi):
(1,2), (2,1), (3,3).
(a) Find the best (in the 2-norm sense) line of the form y = ax + fJ that fits this
data.
(b) Find the best (in the 2-norm sense) line of the form x = ay + fJ that fits this
data.
3. Suppose q, and qz are two orthonormal vectors and b is a fixed vector, all in ffi.
n

(a) Find the optimallinear combination aql + (3q2 that is closest to b (in the 2-norm
sense).
(b) Let r denote the "error vector" b - aql - {3qz. Show that r is orthogonal to
both ql and q2.
Exercises 73
4. Find all solutions of the linear least squares problem
5. Consider the problem of finding the minimum 2-norm solution of the linear least
rmarp nrrh1<=>m
(a) Consider a perturbation E\ = [
0
pi of A, where 8 is a small positive number.
Solve the perturbed version of the above problem,
where AI = A + E\. What happens to ||jt* y ||
2
as 8 approaches 0?
(b) Now consider the perturbation EI = \
0 s
~\ of A, where again 8 is a small
positive number. Solve the perturbed problem
where A
2
A + E
2
. What happens to \\x* z||
2
as 8 approaches 0?
6. Use the four Penrose conditions and the fact that Q\ has orthonormal columns to
verify that if A e R
x
" can be factored in the form (8.9), then A+ = R~
l
Q\.
1. Let A e R"
x
", not necessarily nonsingular, and suppose A = QR, where Q is
orthogonal. Prove that A
+
= R
+
Q
T
.
Exercises 73
4. Find all solutions of the linear least squares problem
min II Ax - bll
2
x
when A = [
5. Consider the problem of finding the minimum 2-norm solution of the linear least
squares problem
min II Ax - bl1
2
x
when A = ] and b = [ ! 1 The solution is
(a) Consider a perturbation EI = of A, where 8 is a small positive number.
Solve the perturbed version of the above problem,
where AI = A + E
I
. What happens to IIx* - yII2 as 8 approaches O?
(b) Now consider the perturbation E2 = n of A, where again 8 is a small
positive number. Solve the perturbed problem
min II A
2
z - bib
z
where A2 = A + E
2
What happens to IIx* - zll2 as 8 approaches O?
6. Use the four Penrose conditions and the fact that QI has orthonormal columns to
verify that if A E can be factored in the form (8.9), then A+ = R-
I
Qf.
7. Let A E not necessarily nonsingular, and suppose A = QR, where Q is
orthogonal. Prove that A + = R+ QT .
Chapter 9
Eigenvalues and
Eigenvectors
Definition 9.1. A nonzero vector x e C" is a right eigenvector of A e C
nxn
if there exists
a scalar A. e C, called an eigenvalue, such that
Similarly, a nonzero vector y e C" is a left eigenvector corresponding to an eigenvalue
a if
By taking Hermitian transposes in (9.1), we see immediately that X
H
is a left eigen-
vector of A
H
associated with A . Note that if x [y] is a right [left] eigenvector of A, then
so is ax [ay] for any nonzero scalar a E C. One often-used scaling for an eigenvector is
a \j'||;t|| so that the scaled eigenvector has norm 1. The 2-norm is the most common
norm used for such scaling.
Definition 9.2. The polynomial n (A.) = det(A A ,/ ) is called the characteristic polynomial
of A. (Note that the characteristic polynomial can also be defined as det(A . / A ). This
results in at most a change of sign and, as a matter of convenience, we use both forms
throughout the text.}
The following classical theorem can be very useful in hand calculation. It can be
proved easily from the Jordan canonical form to be discussed in the text to follow (see, for
example, [21]) or directly using elementary properties of inverses and determinants (see,
for example, [3]).
Theorem 9.3 (Cayley-Hamilton). For any A e C
nxn
, n(A) = 0.
Example 9.4. Let A = [~g ~g] . Then n(k) = X
2
+ 2A , 3. It is an easy exercise to
verify that n(A) = A
2
+ 2A - 31 = 0.
It can be proved from elementary properties of determinants that if A e C"
x
", then
7 t (X) is a polynomial of degree n. Thus, the Fundamental Theorem of A lgebra says that
75
Chapter 9
Eigenvalues and
Eigenvectors
Definition 9.1. A nonzero vector x E en is a right eigenvector of A E e
nxn
if there exists
a scalar A E e, called an eigenvalue, such that
Ax = AX. (9.1)
Similarly, a nonzero vector y E en is a left eigenvector corresponding to an eigenvalue
Mif
(9.2)
By taking Hennitian transposes in (9.1), we see immediately that x
H
is a left eigen-
vector of A H associated with I. Note that if x [y] is a right [left] eigenvector of A, then
so is ax [ay] for any nonzero scalar a E C. One often-used scaling for an eigenvector is
a = 1/ IIx II so that the scaled eigenvector has nonn 1. The 2-nonn is the most common
nonn used for such scaling.
Definition 9.2. The polynomialn (A) = det (A - A l) is called the characteristic polynomial
of A. (Note that the characteristic polynomial can also be defined as det(Al - A). This
results in at most a change of sign and, as a matter of convenience, we use both forms
throughout the text.)
The following classical theorem can be very useful in hand calculation. It can be
proved easily from the Jordan canonical fonn to be discussed in the text to follow (see, for
example, [21D or directly using elementary properties of inverses and determinants (see,
for example, [3]).
Theorem 9.3 (Cayley-Hamilton). For any A E e
nxn
, n(A) = O.
Example 9.4. Let A = [ - ~ - ~ ] . Then n(A) = A2 + 2A - 3. It is an easy exercise to
verify that n(A) = A2 + 2A - 31 = O.
It can be proved from elementary properties of detenninants that if A E e
nxn
, then
n(A) is a polynomial of degree n. Thus, the Fundamental Theorem of Algebra says that
75
and set X = 0 in this identity, we get the interesting fact that del (A) = AI A.2 A
M
(see
also Theorem 9.25).
If A e W
xn
, then n(X) has real coefficients. Hence the roots of 7 r( A) , i.e., the
eigenvalues of A, must occur in complex conjugate pairs.
Example 9.6. Let a, ft e R and let A = [ _^ ]. Then jr( A. ) = A.
2
- 2aA + a
2
+ ft
2
and
A has eigenvalues a fij (where j = i = >/!)
If A R"
x
", then there is an easily checked relationship between the left and right
eigenvectors of A and A
T
(take Hermitian transposes of both sides of (9.2)). Specifically, if
y is a left eigenvector of A corresponding to A e A( A) , then y is a right eigenvector of A
T
corresponding to A. A ( A) . Note, too, that by elementary properties of the determinant,
we always have A ( A ) = A ( A
r
) , but that A ( A ) = A ( A ) only if A e R"
x
".
Definition 9.7. IfX is a root of multiplicity m ofjr(X), we say that X is an eigenvalue of A
of algebraic multiplicity m. The geometric multiplicity ofX is the number of associated
independent eigenvectors = n rank( A A/) = dim J \ f(A XI).
If A A ( A ) has algebraic multiplicity m, then 1 < di mA/ "(A A/) < m. Thus, if
we denote the geometric multiplicity of A by g, then we must have 1 < g < m.
Definition 9.8. A matrix A e W
x
" is said to be defective if it has an eigenvalue whose
geometric multiplicity is not equal to (i.e., less than) its algebraic multiplicity. Equivalently,
A is said to be defective if it does not have n linearly independent (right or left) eigenvectors.
From the Cayley-Hamilton Theorem, we know that n(A) = 0. However, it is pos-
sible for A to satisfy a lower-order polynomial. For example, if A = \
1
Q
], then A sat-
isfies (1 I)
2
= 0. But it also clearly satisfies the smaller degree polynomial equation
a - n = o.
Definition 5.5. The minimal polynomial of A G K""" is the polynomial o/ (X) of least
degree such that a (A) =0.
It can be shown that or(l) is essentially unique (unique if we force the coefficient
of the highest power of A to be +1, say; such a polynomial is said to be monic and we
generally write et (A) as a monic polynomial throughout the text). Moreover, it can also be
7 6 Chapt er 9. Ei g e n va l ue s and Ei genvect ors
7 r( A) has n roots, possibly repeated. These roots, as solutions of the determinant equation
are the eigenvalues of A and imply the singularity of the matrix A XI, and hence further
guarantee the existence of corresponding nonzero eigenvectors.
Definition 9.5. The spectrum of A e C"
x
" is the set of all eigenvalues of A, i.e., the set of
all roots of its characteristic polynomial n(X). The spectrum of A is denoted A ( A) .
Let the eigenvalues of A e C"
x
" be denoted X\ ,..., X
n
. Then if we write (9.3) in the
form
76 Chapter 9. Eigenvalues and Eigenvectors
n(A) has n roots, possibly repeated. These roots, as solutions of the determinant equation
n(A) = det(A - AI) = 0, (9.3)
are the eigenvalues of A and imply the singularity of the matrix A - AI, and hence further
guarantee the existence of corresponding nonzero eigenvectors.
Definition 9.5. The spectrum of A E c
nxn
is the set of all eigenvalues of A, i.e., the set of
all roots of its characteristic polynomialn(A). The spectrum of A is denoted A(A).
Let the eigenvalues of A E en xn be denoted A], ... , An. Then if we write (9.3) in the
form
n(A) = det(A - AI) = (A] - A) ... (An - A) (9.4)
and set A = 0 in this identity, we get the interesting fact that det(A) = A] . A2 ... An (see
also Theorem 9.25).
If A E 1Ftnxn, then n(A) has real coefficients. Hence the roots of n(A), i.e., the
eigenvalues of A, must occur in complex conjugate pairs.
Example 9.6. Let a, f3 E 1Ft and let A = ~ f 3 !]. Then n(A) = A
2
- 2aA + a
2
+ f32 and
A has eigenvalues a f3j (where j = i = R).
If A E 1Ftnxn, then there is an easily checked relationship between the left and right
eigenvectors of A and AT (take Hermitian transposes of both sides of (9.2. Specifically, if
y is a left eigenvector of A corresponding to A E A(A), then y is a right eigenvector of AT
corresponding to I E A(A). Note, too, that by elementary properties of the determinant,
we always have A(A) = A(AT), but that A(A) = A(A) only if A E 1Ftnxn.
Definition 9.7. If A is a root of multiplicity m of n(A), we say that A is an eigenvalue of A
of algebraic multiplicity m. The geometric multiplicity of A is the number of associated
independent eigenvectors = n - rank(A - AI) = dimN(A - AI).
If A E A(A) has algebraic multiplicity m, then I :::: dimN(A - AI) :::: m. Thus, if
we denote the geometric multiplicity of A by g, then we must have I :::: g :::: m.
Definition 9.8. A matrix A E 1Ft
nxn
is said to be defective if it has an eigenvalue whose
geometric multiplicity is not equal to (i.e., less than) its algebraic multiplicity. Equivalently,
A is said to be defective if it does not have n linearly independent (right or left) eigenvectors.
From the Cayley-Hamilton Theorem, we know that n(A) = O. However, it is pos-
sible for A to satisfy a lower-order polynomial. For example, if A = [ ~ ~ ] , then A sat-
isfies (Je - 1)2 = O. But it also clearly satisfies the smaller degree polynomial equation
(it. - 1) ;;;:; 0
neftnhion ~ . ~ . Thll minimal polynomial Of A l::: l!if.nxn ix (hI' polynomilll a(A) oJ IPll.ft
degree such that a(A) ~ O.
It can be shown that a(Je) is essentially unique (unique if we force the coefficient
of the highest power of A to be + 1. say; such a polynomial is said to be monic and we
generally write a(A) as a monic polynomial throughout the text). Moreover, it can also be
9.1. Fundamental Definitions and Properties 77
shown that a (A.) divides every nonzero polynomial fi(k} for which ft (A) = 0. In particular,
a(X) divides n(X).
There is an algorithm to determine or ( A . ) directly ( without knowing eigenvalues and as-
sociated eigenvector structure). Unfortunately, this algorithm, called the Bezout algorithm,
is numerically unstable.
Example 9.10. The above definitions are illustrated below for a series of matrices, each
of which has an eigenvalue 2 of algebraic multiplicity 4, i. e. , 7r( A ) = ( A 2)
4
. We denote
the geometric multiplicity by g.
A t this point, one might speculate that g plus the degree of a must always be five.
Unfortunately, such is not the case. The matrix
Theorem 9.11. Let A e C
x
"
ana
[
e
t A ., be an eigenvalue of A with corresponding right
eigenvector j c,-. Furthermore, let yj be a left eigenvector corresponding to any A
;
e A ( A )
such that Xj = A . ,. Then yfx{ = 0.
Proof: Since Ax
t
= A ,*,,
shown that a(A) divides every nonzero polynomial f3(A) for which f3(A) = O. In particular,
a(A) divides n(A).
There is an algorithm to determine a(A) directly (without knowing eigenvalues and as-
sociated eigenvector structure). Unfortunately, this algorithm, called the Bezout algorithm,
is numerically unstable.
Example 9.10. The above definitions are illustrated below for a series of matrices, each
of which has an eigenvalue 2 of algebraic multiplicity 4, i.e., n(A) = (A - 2)4. We denote
the geometric multiplicity by g.
A - [ ~
0
! ] ha,"(A) ~ (A - 2)' ""d g ~ 1.
2 I
- 0
0 2
0 0 0
A ~ [ ~
0
~ ] ha< a(A) ~ (A - 2)' ""d g ~ 2.
2
0 2
0 0
A ~ U
I 0
~ ] h'" a(A) ~ (A - 2)2 ""d g ~ 3.
2 0
0 2
0 0
A ~ U
0 0
~ ] ha<a(A) ~ (A - 2) andg ~ 4.
2 0
0 2
0 0
At this point, one might speculate that g plus the degree of a must always be five.
Unfortunately, such is not the case. The matrix
A ~ U
I 0
!]
2 0
0 2
0 0
has a(A) = (A - 2)2 and g = 2.
Theorem 9.11. Let A E cc
nxn
and let Ai be an eigenvalue of A with corresponding right
eigenvector Xi. Furthermore, let Yj be a left eigenvector corresponding to any Aj E l\(A)
such that Aj 1= Ai. Then YY Xi = O.
Proof' Since AXi = AiXi,
(9.5)
Similarly, since y" A = Xjyf,
Subtracting (9.6) from (9.5), we find 0 = (A.,- A
y
)j ^j c, . Since A,,- A.
7
- ^ 0, we must have
yfxt =0.
The proof of Theorem 9.11 is very similar to two other fundamental and important
results.
Theorem 9.12. Let A e C"
x
" be Hermitian, i.e., A = A
H
. Then all eigenvalues of A must
be real.
Proof: Suppose (A ., x) is an arbitrary eigenvalue/eigenvector pair such that Ax = A .J C. Then
Taking Hermitian transposes in (9.7) yields
Using the fact that A is Hermitian, we have that Xx
H
x = Xx
H
x. However, since x is an
eigenvector, we have X
H
X /= 0, from which we conclude A . = A , i.e., A . is real. D
x
" be Hermitian and suppose X and / J L are distinct eigenvalues
of A with corresponding right eigenvectors x and z, respectively. Then x and z must be
orthogonal.
Proof: Premultiply the equation Ax = A.J C by Z
H
to get Z
H
Ax = X z
H
x . Take the Hermitian
transpose of this equation and use the facts that A is Hermitian and A . is real to get X
H
Az =
Xx
H
z. Premultiply the equation Az = i^z by X
H
to get X
H
Az = / ^X
H
Z = Xx
H
z. Since
A, ^ /z, we must have that X
H
z = 0, i.e., the two vectors must be orthogonal. D
Let us now return to the general case.
Theorem 9.14. Let A . C
nxn
have distinct eigenvalues A ,
1 ?
. . . , A .
n
with corresponding
right eigenvectors x\,... ,x
n
. Then [x\,..., x
n
} is a linearly independent set. The same
result holds for the corresponding left eigenvectors.
Proof: For the proof see, for example, [21, p. 118].
If A e C
nx
" has distinct eigenvalues, and if A ., e A (A ), then by Theorem 9.11, jc, is
orthogonal to all yj's for which j ^ i. However, it cannot be the case that yf*x
t
= 0 as
well, or else x
f
would be orthogonal to n linearly independent vectors (by Theorem 9.14)
and would thus have to be 0, contradicting the fact that it is an eigenvector. Since yf*Xi ^ 0
for each i, we can choose the normalization of the *, 's, or the y, 's, or both, so that y
t
H
x; = 1
f or / n.
Similarly, since YY A = A j yy,
(9.6)
Subtracting (9.6) from (9.5), we find 0 = (Ai - Aj)YY xi. Since Ai - Aj =1= 0, we must have
YyXi = O. 0
The proof of Theorem 9.11 is very similar to two other fundamental and important
results.
Theorem 9.12. Let A E c
nxn
be Hermitian, i.e., A = AH. Then all eigenvalues of A must
be real.
Proof: Suppose (A, x) is an arbitrary eigenvalue/eigenvector pair such that Ax = AX. Then
(9.7)
Taking Hermitian transposes in (9.7) yields
Using the fact that A is Hermitian, we have that IXH x = AXH x. However, since x is an
eigenvector, we have xH x =1= 0, from which we conclude I = A, i.e., A is real. 0
nxn
be Hermitian and suppose A and iJ- are distinct eigenvalues
of A with corresponding right eigenvectors x and z, respectively. Then x and z must be
orthogonal.
Proof: Premultiply the equation Ax = AX by ZH to get ZH Ax = AZ
H
x. Take the Hermitian
transpose of this equation and use the facts that A is Hermitian and A is real to get x H Az =
AxH z. Premultiply the equation Az = iJ-Z by x
H
to get x
H
Az = iJ-XH Z = AXH z. Since
A =1= iJ-, we must have that x
H
z = 0, i.e., the two vectors must be orthogonal. 0
Let us now return to the general case.
nxn
have distinct eigenvalues AI, ... , An with corresponding
right eigenvectors XI, ... , x
n
Then {XI, ... , x
n
} is a linearly independent set. The same
result holds for the corresponding left eigenvectors.
Proof: For the proof see, for example, [21, p. 118]. 0
If A E c
nxn
has distinct eigenvalues, and if Ai E A(A), then by Theorem 9.11, Xi is
orthogonal to all y/s for which j =1= i. However, it cannot be the case that Yi
H
Xi = 0 as
well, or else Xi would be orthogonal to n linearly independent vectors (by Theorem 9.14)
and would thus have to be 0, contradicting the fact that it is an eigenvector. Since yr Xi =1= 0
for each i, we can choose the normalization of the Xi'S, or the Yi 's, or both, so that Yi
H
Xi = 1
for i E !1.
9.1. Fundament al Def i ni t i o ns and Properties 79
x
" have distinct eigenvalues A .I , ..., A .
n
and let the correspond-
ing right eigenvectors form a matrix X = [x\, ..., x
n
]. Similarly, let Y [y\, ..., y
n
]
be the matrix of corresponding left eigenvectors. Furthermore, suppose that the left and
right eigenvectors have been normalized so that yf
1
Xi = 1, / en. Finally, let A =
di ag ( A ,j , . . . , X
n
) e W
txn
. Then A J C, = A .,- * /, / e n, can be written in matrix form as
Example 9.16. Let
Then n(X) = det( A - A ./) = - (A .
3
+ 4A .
2
+ 9 A . + 10) = - (A . + 2 )(A .
2
+ 2 A , + 5), from
which we find A ( A ) = { 2 , 1 2 j } . We can now find the right and left eigenvectors
corresponding to these eigenvalues.
For A - i = 2 , solve the 3 x 3 linear system (A (2 } I)x\ = 0 to get
while y^Xj = 5,
;
, / en, y' e n, is expressed by the equation
These matrix equations can be combined to yield the following matrix factorizations:
and
Note that one component of ;ci can be set arbitrarily, and this then determines the other two
(since di mA /XA ( 2 )7) = 1). To get the corresponding left eigenvector y\, solve the
linear system y\(A + 2 1) = 0 to get
This time we have chosen the arbitrary scale factor for y\ so that y f x \ = 1.
For A
2
= 1 + 2 j , solve the linear system (A ( 1 + 2 j )I)x
2
= 0 to get
Theorem 9.15. Let A E en xn have distinct eigenvalues A I, ... , An and let the correspond-
ing right eigenvectors form a matrix X = [XI, ... , xn]. Similarly, let Y = [YI,"" Yn]
be the matrix of corresponding left eigenvectors. Furthermore, suppose that the left and
right eigenvectors have been normalized so that YiH Xi = 1, i E !!:: Finally, let A =
diag(AJ, ... , An) E ]Rnxn. Then AXi = AiXi, i E !!, can be written in matrixform as
AX=XA (9.8)
while YiH X j = oij, i E!!, j E !!, is expressed by the equation
yHX = I.
(9.9)
These matrix equations can be combined to yield the following matrix factorizations:
and
Example 9.16. Let
X-lAX = A = yRAX
n
A = XAX-
I
= XAyH = LAixiyr
2
5
-3
-3
-2
i=1
~ ] .
-4
(9.10)
(9.11)
Then rr(A) = det(A - AI) = -(A
3
+ 4A2 + 9)" + 10) = -()" + 2)(),,2 + 2)" + 5), from
which we find A(A) = {-2, -1 2j}. We can now find the right and left eigenvectors
corresponding to these eigenvalues.
For Al = -2, solve the 3 x 3 linear system (A - (-2)l)xI = 0 to get
Note that one component of XI can be set arbitrarily, and this then determines the other two
(since dimN(A - (-2)1) = 1). To get the corresponding left eigenvector YI, solve the
linear system yi (A + 21) = 0 to get
This time we have chosen the arbitrary scale factor for YJ so that yi XI = 1.
For A2 = -1 + 2j, solve the linear system (A - (-1 + 2j) I)x2 = 0 to get
[
3+ j ]
X2 = 3 / .
Solve the linear system y" (A (-1 + 27')/) = 0 and normalize y>
2
so that y"x
2
= 1 to get
For X T , = 1 2 j, we could proceed to solve linear systems as for A.
2
. However, we
can also note that x$ =x
2
' and yi = jj. To see this, use the fact that A, 3 = A.2 and simply
conjugate the equation A;c
2
^.2 *2 to get Ax^ = ^2 X 2 - A similar argument yields the result
for left eigenvectors.
Now define the matrix X of right eigenvectors:
It is then easy to verify that
Other results in Theorem 9.15 can also be verified. For example,
Finally, note that we could have solved directly only for *i and x
2
(and X T , = x
2
). Then,
instead of determining the j,'s directly, we could have found them instead by computing
X ~
l
and reading off its rows.
Example 9.17. Let
Then 7r(A.) = det(A - A./) = -(A
3
+ 8A
2
+ 19X + 12) = -(A. + 1)(A. + 3)(A, + 4),
from which we find A (A) = { 1, 3, 4}. Proceeding as in the previous example, it is
straightforward to compute
and
Solve the linear system yf (A - ( -I + 2 j) I) = 0 and nonnalize Y2 so that yf X2 = 1 to get
For A3 = -I - 2j, we could proceed to solve linear systems as for A2. However, we
can also note that X3 = X2 and Y3 = Y2. To see this, use the fact that A3 = A2 and simply
conjugate the equation AX2 = A2X2 to get AX2 = A2X2. A similar argument yields the result
for left eigenvectors.
Now define the matrix X of right eigenvectors:
3+j 3-
j
]
3-j 3+j .
-2 -2
It is then easy to verify that
.!.=.L
4
l+j
4
!.1
4
.!.=.L
4
Other results in Theorem 9.15 can also be verified. For example,
[
-2 0
X-IAX=A= 0 -1+2j
o 0
Finally, note that we could have solved directly only for XI and X2 (and X3 = X2). Then,
instead of detennining the Yi'S directly, we could have found them instead by computing
X-I and reading off its rows.
Example 9.17. Let
A = .
o -3
Then Jl"(A) = det(A - AI) = _(A
3
+ 8A
2
+ 19A + 12) = -(A + I)(A + 3)(A + 4),
from which we find A(A) = {-I, -3, -4}. Proceeding as in the previous example, it is
gtruightforw!U"d to

I
-i ]
0
-I
and

1 2 1
] y'
3 0 -3
2 -2 2
We also have X~
l
AX = A = di ag( 1, 3, 4 ) , which is equivalent to the dyadic expan-
sion
Theorem 9.18. Eigenvalues (but not eigenvectors) are invariant under a similarity trans-
formation T.
Proof: Suppose (A, jc) is an eigenvalue/eigenvector pair such that Ax = Xx. Then, since T
is nonsingular, we have the equivalent statement (T~
l
AT)(T~
l
x) = X ( T ~
l
x ) , from which
the theorem statement follows. For left eigenvectors we have a similar statement, namely
y
H
A = Xy
H
ifandon\yif(T
H
y)
H
(T~
1
AT) =X(T
H
yf. D
Remark 9.19. If / is an analytic function (e.g., f ( x ) is a polynomial, or e
x
, or sin*,
or, in general, representable by a power series X^ô
fl
n*
n
)> then it is easy to show that
the eigenvalues of /( A) (defined as Xô^-A") are /( A) , but /( A) does not necessarily
have all the same eigenvectors (unless, say, A is diagonalizable). For example, A = T
0 O
j
has only one right eigenvector corresponding to the eigenvalue 0, but A
2
= f
0 0
1 has two
independent right eigenvectors associated with the eigenvalue 0. What is true is that the
eigenvalue/eigenvector pair (A, x) maps to ( /( A) , x) but not conversely.
The following theorem is useful when solving systems of linear differential equations.
Details of how the matrix exponential e'
A
is used to solve the system x = Ax are the subject
of Chapter 11.
Theorem 9.20. Let A e R"
xn
and suppose X~~
1
AX A, where A is diagonal. Then
We also have X-I AX = A = diag( -1, -3, -4), which is equivalent to the dyadic expan-
sion
3
A = LAixiyr
i=1
W j 0
+(-4) [ -; ]
1

- -
3
(-I) [
I I I
J + (-3) [
I
0
I
] + (-4) [
I I I
l
(;
3
(;
2
-2 3
-3
3
I 2 I
0 0 0
I I I
3 3 3
-3
3
-3
I I I
I
0
I
I I I
(;
3
(;
-2
2
3
-3
3
Theorem 9.18. Eigenvalues (but not eigenvectors) are invariant under a similarity trans-
formation T.
Proof: Suppose (A, X) is an eigenvalue/eigenvector pair such that Ax = AX. Then, since T
is nonsingular, we have the equivalent statement (T-
I
AT)(T-
I
x) = A(T-
I
x), from which
the theorem statement follows. For left eigenvectors we have a similar statement, namely
yH A = AyH if and only if (T
H
y)H (T-
1
AT) = A(T
H
y)H. D
Remark 9.19. If f is an analytic function (e.g., f(x) is a polynomial, or eX, or sinx,
or, in general, representable by a power series anxn), then it is easy to show that
the eigenvalues of f(A) (defined as are f(A), but f(A) does not necessarily
have all the same eigenvectors (unless, say, A is diagonalizable). For example, A = 6 ]
has only one right eigenvector corresponding to the eigenvalue 0, but A
2
= ] has two
independent right eigenvectors associated with the eigenvalue o. What is true is that the
eigenvalue/eigenvector pair (A, x) maps to (f(A), x) but not conversely.
The following theorem is useful when solving systems of linear differential equations.
Details of how the matrix exponential etA is used to solve the system i = Ax are the subject
of Chapter 11.
Theorem 9.20. Let A E jRnxn and suppose X-I AX = A, where A is diagonal. Then
n
= LeA,txiYiH.
i=1
Proof: Starting from the definition, we have
The following corollary is immediate from the theorem upon setting t = I.
Corollary 9.21. If A e R
nx
" is diagonalizable with eigenvalues A .,- , /' en, and right
eigenvectors x
t
, / n_, then e
A
has eigenvalues e
X i
, i n_, and the same eigenvectors.
There are extensions to Theorem 9.20 and Corollary 9.21 for any function that is
analytic on the spectrum of A , i.e., f ( A ) = X f ( A ) X ~
l
= Xdi ag ( / ( A . i ) , . . . , f ( X
t t
) ) X ~
l
.
It is desirable, of course, to have a version of Theorem 9.20 and its corollary in which
A is not necessarily diagonalizable. It is necessary first to consider the notion of Jordan
canonical form, from which such a result is then available and presented later in this chapter.
9.2 Jordan Canonical Form
Theorem 9.22.
1. Jordan Canonical Form (/CF): For all A e C"
x
" with eigenvalues X\,..., k
n
e C
(not necessarily distinct), there exists X C^
x
" such that
where each of the Jordan block matrices / i , . . . , J
q
is of the form
Proof' Starting from the definition, we have
n
= LeA;IXiYiH. 0
i=1
The following corollary is immediate from the theorem upon setting t = I.
Corollary 9.21. If A E ]Rn xn is diagonalizable with eigenvalues Ai, i E ~ , and right
eigenvectors Xi, i E ~ , then e
A
has eigenvalues e
A
" i E ~ , and the same eigenvectors.
There are extensions to Theorem 9.20 and Corollary 9.21 for any function that is
analytic on the spectrum of A, i.e., f(A) = Xf(A)X-
I
= Xdiag(J(AI), ... , f(An))X-
I
.
It is desirable, of course, to have a version of Theorem 9.20 and its corollary in which
A is not necessarily diagonalizable. It is necessary first to consider the notion of Jordan
canonical form, from which such a result is then available and presented later in this chapter.
9.2 Jordan Canonical Form
Theorem 9.22.
I. lordan Canonical Form (JCF): For all A E c
nxn
with eigenvalues AI, ... , An E C
(not necessarily distinct), there exists X E c ~ x n such that
X-I AX = 1 = diag(ll, ... , 1q), (9.12)
where each of the lordan block matrices 1
1
, , 1q is of the form
Ai
0 o
0
Ai
0
1i =
Ai
(9.13)
o
Ai
o o Ai
9.2. Jordan Canonical Form 83
2. Real Jordan Canonical Form: For all A R
nx
" with eigenvalues Xi, . . . , X
n
(not
necessarily distinct), there exists X R"
xn
such that
where each of the Jordan block matrices J\, ..., J
q
is of the form
in the case of real eigenvalues A., e A (A), and
where M
;
= [ _' ^ 1 and I
2
= \
0
A in the case of complex conjugate eigenvalues
a
i
jp
i
eA(A
>
).
Proof: For the proof see, for example, [21, pp. 120-124]. D
Transformations like T = I " _, "{ "] allow us to go back and forth between a real JCF
and its complex counterpart:
For nontrivial Jordan blocks, the situation is only a bit more complicated. With
9.2. Jordan Canonical Form 83
and L;=1 ki = n.
2. Real Jordan Canonical Form: For all A E jRnxn with eigenvalues AI, ... , An (not
necessarily distinct), there exists X E such that
(9.14)
where each of the Jordan block matrices 11, ... , 1q is of the form
where Mi = [ ] and h = [6 in the case of complex conjugate eigenvalues
(Xi jfJi E A(A).
Proof: For the proof see, for example, [21, pp. 120-124]. 0
Transformations like T = [ _ - { ] allow us to go back and forth between a real JCF
and its complex counterpart:
T-I [ (X + jfJ O. ] T = [ (X fJ ] = M.
o (X - JfJ -fJ (X
For nontrivial Jordan blocks, the situation is only a bit more complicated. With
1
o
-j
o
-j
o
1 o 0 '

o -j 1
84 Chapter 9. Ei genval ues and Eigenvectors
it is easily checked that
Definition 9.23. The characteristic polynomials of the Jordan blocks defined in Theorem
9 . 2 2 are called the elementary divisors or invariant factors of A.
Theorem 9.24. The characteristic polynomial of a matrix is the product of its elementary
divisors. The minimal polynomial of a matrix is the product of the elementary divisors of
highest degree corresponding to distinct eigenvalues.
x
" with eigenvalues AI, . . . , X
n
. Then
Proof:
1. From Theorem 9.22 we have that A = X J X ~
l
. Thus,
det(A) = det(XJX-
1
) = det(7) = ] ~ [ "
=l
A, - .
2. Again, from Theorem 9.22 we have that A = X J X ~
l
. Thus,
Tr(A) = Tr(XJX~
l
) = TrC/X"
1
*) = Tr(/) = "
=1
A.,- . D
Example 9.26. Suppose A e E
7x7
is known to have 7r(A) = (A. - 1)
4
(A - 2)
3
and
a (A.) = (A. I)
2
(A. 2)
2
. Then A has two possible JCFs (not counting reorderings of the
diagonal blocks):
Note that 7
(1)
has elementary divisors (A - I )
2
, (A. - 1), (A. - 1), (A, - 2)
2
, and (A - 2),
while /(
2)
has elementary divisors (A - I )
2
, (A - I )
2
, (A - 2)
2
, and (A - 2).
it is easily checked that
[ "+ jfi
0 0
] T ~ [ ~
T-
I
0
et + jf3 0 0
h
l
0 0 et - jf3 M
0 0 0 et - jf3
Definition 9.23. The characteristic polynomials of the Jordan blocks defined in Theorem
9.22 are called the elementary divisors or invariant factors of A.
Theorem 9.24. The characteristic polynomial of a matrix is the product of its elementary
divisors. The minimal polynomial of a matrix is the product of the elementary divisors of
highest degree corresponding to distinct eigenvalues.
nxn
with eigenvalues AI, .. " An. Then
n
1. det(A) = nAi.
i=1
n
2. Tr(A) = 2,)i.
i=1
Proof:
1. From Theorem 9.22 we have that A = X J X-I. Thus,
det(A) = det(X J X-I) = det(J) = n7=1 Ai.
2. Again, from Theorem 9.22 we have that A = X J X-I. Thus,
Tr(A) = Tr(X J X-I) = Tr(JX-
1
X) = Tr(J) = L7=1 Ai. 0
Example 9.26. Suppose A E lR.
7x7
is known to have :rr(A) = (A - 1)4(A - 2)3 and
et(A) = (A - 1)2(A - 2)2. Then A has two possible JCFs (not counting reorderings of the
diagonal blocks):
1 0 0 0 0 0 1 0 0 0 0 0
0 1 0 0 0 0 0 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0 1 I 0 0 0
J(l) =
0 0 0 1 0 0 0
and f2) =
0 0 0 1 0 0 0
0 0 0 0 2 1 0 0 0 0 0 2 1 0
0 0 0 0 0 2 0 0 0 0 0 0 2 0
0 0 0 0 0 0 2
0 0 0 0 0 0 2
Note that J(l) has elementary divisors (A - 1)z, (A - 1), (A - 1), (A - 2)2, and (A - 2),
while J(2) has elementary divisors (A - 1)2, (A - 1)2, (A - 2)2, and (A - 2).
9.3. Determination of the JCF &5
Example 9.27. Knowing T T (A.), a ( A ) , and rank (A A,,7) for distinct A., is not sufficient to
determine the JCF of A uniquely. T he matrices
both have 7r( A . ) = (A. a) , a( A . ) = (A. a) , and rank( A al) = 4, i.e., three eigen-
vectors.
9.3 Determination of the JCF
T he first critical item of information in determining the JCF of a matrix A e W
lxn
is its
number of eigenvectors. For each distinct eigenvalue A , , , the associated number of linearly
independent right (or left) eigenvectors is given by dim A^(A A.,7) = n rank( A A.
(
7).
T he straightforward case is, of course, when X,- is simple, i.e., of algebraic multiplicity 1; it
then has precisely one eigenvector. T he more interesting (and difficult) case occurs when
A, is of algebraic multiplicity greater than one. For example, suppose
T hen
has rank 1, so the eigenvalue 3 has two eigenvectors associated with it. If we let [î 2 &]
T
denote a solution to the linear system (A 3/) = 0, we find that 2
2
+ 3=0. T hus, both
are eigenvectors (and are independent). T o get a third vector JC3 such that X = [x\ KJ_ XT,]
reduces A to JCF, we need the notion of principal vector.
Definition 9.28. Let A e C"
xn
(or R"
x
") . Then x is a right principal vector of degree k
associated with X e A (A) if and only if (A - XI)
k
x = 0 and (A - U}
k
~
l
x ^ 0.
Remark 9.29.
1. An analogous definition holds for a left principal vector of degree k.
9.3. Determination of the JCF 85
Example 9.27. Knowing rr(A), a(A), and rank(A - Ai l) for distinct Ai is not sufficient to
determine the JCF of A uniquely. The matrices
a 0 0 0 0 0 a 0 0 0 0 0
0 a 0 0 0 0 0 a 0 0 0 0
0 0 a 0 0 0 0 0 0 a 0 0 0 0
Al=
0 0 0 a 0 0
A2 =
0 0 0 a 0 0
0 0 0 0 a 0 0 0 0 0 0 a 0
0 0 0 0 0 a 1 0 0 0 0 0 a 0
0 0 0 0 0 0 a 0 0 0 0 0 0 a
both have rr(A) = (A - a)7, a(A) = (A - a)\ and rank(A - al) = 4, i.e., three eigen-
vectors.
9.3 Determination of the JCF
The first critical item of information in determining the JCF of a matrix A E ]R.nxn is its
number of eigenvectors. For each distinct eigenvalue Ai, the associated number of linearly
independent right (or left) eigenvectors is given by dimN(A - A;l) = n - rank(A - A;l).
The straightforward case is, of course, when Ai is simple, i.e., of algebraic multiplicity 1; it
then has precisely one eigenvector. The more interesting (and difficult) case occurs when
Ai is of algebraic multiplicity greater than one. For example, suppose
[3 2
n
A = 0 3
o 0
Then
A-3I= U
2 I]
o 0
o 0
has rank 1, so the eigenvalue 3 has two eigenvectors associated with it. If we let
denote a solution to the linear system (A - = 0, we find that + = O. Thus, both
are eigenvectors (and are independent). To get a third vector X3 such that X = [Xl X2 X3]
reduces A to JCF, we need the notion of principal vector.
Definition 9.28. Let A E c
nxn
(or ]R.nxn). Then X is a right principal vector of degree k
associated with A E A(A) ifand only if(A - ulx = 0 and (A - AI)k-l x i= o.
Remark 9.29.
1. An analogous definition holds for a left principal vector of degree k.
2. The phrase "of grade k" is often used synonymously with "of degree k."
3. Principal vectors are sometimes also called generalized eigenvectors, but the latter
term will be assigned a much different meaning in Chapter 12.
4. The case k = 1 corresponds to the "usual" eigenvector.
5. A right (or left) principal vector of degree k is associated with a Jordan block ji of
dimension k or larger.
9.3.1 Theoretical computation
To motivate the development of a procedure for determining principal vectors, consider a
2x2 Jordan block {h
0
h1. Denote by x
(1)
and x
(2)
the two columns of a matrix X e R
2
,x
2
that reduces a matrix A to this JCF. Then the equation AX = XJ can be written
The first column yields the equation Ax
(1)
= hx
(1)
which simply says that x
(1)
is a right
eigenvector. The second column yields the following equation for x
(2)
, the principal vector
of degree 2:
If we premultiply (9.17) by (A - XI), we find (A - XI )
z
x
( 2 )
= (A - XI)x
w
= 0. Thus,
the definition of principal vector is satisfied.
This suggests a "general" procedure. First, determine all eigenvalues of A e R"
x
"
(or C
nxn
). Then for each distinct X e A (A) perform the following:
1. Solve
This step finds all the eigenvectors (i.e., principal vectors of degree 1) associated with
X. The number of eigenvectors depends on the rank of A XI. For example, if
rank(A XI) = n 1, there is only one eigenvector. If the algebraic multiplicity of
X is greater than its geometric multiplicity, principal vectors still need to be computed
from succeeding steps.
2. For each independent jc
(1)
, solve
The number of linearly independent solutions at this step depends on the rank of
(A XI )
2
. If, for example, this rank is n 2 , there are two linearly independent
solutions to the homogeneous equation (A XI)
2
x^ = 0. One of these solutions
is, of course, x
(l)
(^ 0), since (A - XI )
2
x
( l )
= (A - XI)0 = 0. The other solution
is the desired principal vector of degree 2. (It may be necessary to take a linear
combination of jc
(1)
vectors to get a right-hand side that is in 7(A XI). See, for
example, Exercise 7.)
2. The phrase "of grade k" is often used synonymously with "of degree k."
3. Principal vectors are sometimes also called generalized eigenvectors, but the latter
term will be assigned a much different meaning in Chapter 12.
4. The case k = 1 corresponds to the "usual" eigenvector.
S. A right (or left) principal vector of degree k is associated with a Jordan block J; of
dimension k or larger.
9.3.1 Theoretical computation
To motivate the development of a procedure for determining principal vectors, consider a
2 x 2 Jordan block [ ~ i]. Denote by x(l) and x(2) the two columns of a matrix X E l R ~ X 2
that reduces a matrix A to this JCF. Then the equation AX = X J can be written
A [x(l) x(2)] = [x(l) X(2)] [ ~ ~ J.
The first column yields the equation Ax(!) = AX(!), which simply says that x(!) is a right
eigenvector. The second column yields the following equation for x(2), the principal vector
of degree 2:
(A - A/)x(2) = x(l). (9.17)
If we premultiply (9.17) by (A - AI), we find (A - A1)2 x(2) = (A - A1)X(l) = O. Thus,
the definition of principal vector is satisfied.
This suggests a "general" procedure. First, determine all eigenvalues of A E lR
nxn
(or c
nxn
). Then for each distinct A E A(A) perform the following:
1. Solve
(A - A1)X(l) = O.
This step finds all the eigenvectors (i.e., principal vectors of degree I) associated with
A. The number of eigenvectors depends on the rank of A - AI. For example, if
rank(A - A/) = n - 1, there is only one eigenvector. If the algebraic multiplicity of
A is greater than its geometric multiplicity, principal vectors still need to be computed
from succeeding steps.
2. For each independent x(l), solve
(A - A1)x(2) = x(l).
The number of linearly independent solutions at this step depends on the rank of
(A - uf. If, for example, this rank is n - 2, there are two linearly independent
solutions to the homogeneous equation (A - AI)2x (2) = o. One of these solutions
is, of course, x(l) (1= 0), since (A - 'A1)
2
x(l) = (A - AI)O = o. The other solution
is the desired principal vector of degree 2. (It may be necessary to take a linear
combination of x(l) vectors to get a right-hand side that is in R(A - AI). See, for
example, Exercise 7.)
9. 3. Determination of the JCF 87
4. Continue in this way until the total number of independent eigenvectors and principal
vectors is equal to the algebraic multiplicity of A.
Unfortunately, this natural-looking procedure can fail to find all Jordan vectors. For
more extensive treatments, see, for example, [20] and [21]. Determination of eigenvectors
and principal vectors is obviously very tedious for anything beyond simple problems (n = 2
or 3, say). Attempts to do such calculations in finite-precision floating-point arithmetic
generally prove unreliable. There are significant numerical difficulties inherent in attempting
to compute a JCF, and the interested student is strongly urged to consult the classical and very
readable [8] to learn why. Notice that high-quality mathematical software such as MATLAB
does not offer a jcf command, although a jordan command is available in MATLAB'S
Symbolic Toolbox.
Theorem 9.30. Suppose A e C
kxk
has an eigenvalue A, of algebraic multiplicity k and
suppose further that rank(A AI) = k 1. Let X = [ x
( l )
, . . . , x
(k)
], where the chain of
vectors x(i) is constructed as above. Then
Theorem 9.31. (x
( 1)
, . . . , x
(k)
} is a linearly independent set.
Theorem 9.32. Principal vectors associated with different Jordan blocks are linearly inde-
pendent.
Example 9.33. Let
The eigenvalues of A are A1 = 1, h2 = 1, and h
3
= 2. First, find the eigenvectors associated
with the distinct eigenvalues 1 and 2.
,(1)
(A - 2/)x3(1) = 0 yields
3. For each independent x
(2)
from step 2, solve
9.3. Determination of the JCF 87
3. For each independent X(2) from step 2, solve
(A - AI)x(3) = x(2).
4. Continue in this way until the total number of independent eigenvectors and principal
vectors is equal to the algebraic multiplicity of A.
Unfortunately, this natural-looking procedure can fail to find all Jordan vectors. For
more extensive treatments, see, for example, [20] and [21]. Determination of eigenvectors
and principal vectors is obviously very tedious for anything beyond simple problems (n = 2
or 3, say). Attempts to do such calculations in finite-precision floating-point arithmetic
generally prove unreliable. There are significant numerical difficulties inherent in attempting
to compute a JCF, and the interested student is strongly urged to consult the classical and very
readable [8] to learn why. Notice that high-quality mathematical software such as MATLAB
does not offer a j cf command, although a j ardan command is available in MATLAB's
Symbolic Toolbox.
Theorem 9.30. Suppose A E C
kxk
has an eigenvalue A of algebraic multiplicity k and
suppose further that rank(A - AI) = k - 1. Let X = [x(l), ... , X(k)], where the chain of
vectors x(i) is constructed as above. Then
Theorem 9.31. {x(l), ... , X(k)} is a linearly independent set.
Theorem 9.32. Principal vectors associated with different Jordan blocks are linearly inde-
pendent.
Example 9.33. Let
1 ;].
002
The eigenvalues of A are AI = I, A2 = 1, and A3 = 2. First, find the eigenvectors associated
with the distinct eigenvalues 1 and 2.
(A - = 0 yields
(A- l/)x,
(1)
=0 yields
Then it is easy to check that
9.3.2 On the +1 's in JCF blocks
In this subsection we show that the nonzero superdiagonal elements of a JCF need not be
1 's but can be arbitrary so long as they are nonzero. For the sake of defmiteness, we
consider below the case of a single Jordan block, but the result clearly holds for any JCF.
Supposed A R
nxn
and
Let D = d i a g ( d 1 , . . . , d
n
) be a nonsingular "scaling" matrix. Then
To find a principal vector of degree 2 associated with the multiple eigenvalue 1, solve
( A l/)x,
(2)
= x,
(1)
toeet
Now let
(A - 11)x?J = 0 yields
To find a principal vector of degree 2 associated with the multiple eigenvalue 1, solve
(A - 1I)xl
2
) = xiI) to get
[ 0 ]
(2)
x, = ~ .
Now let
xl" xl"] ~ [ ~
0 5
l
X = [xiI) 1 3
0
Then it is easy to check that
X - ' ~ U
0
-5 ] [ I
n
1
-i and X-lAX = ~
1
0 0
9.3.2 On the +1 's in JCF blocks
In this subsection we show that the nonzero superdiagonal elements of a JCF need not be
1 's but can be arbitrary - so long as they are nonzero. For the sake of definiteness, we
consider below the case of a single Jordan block, but the result clearly holds for any JCF.
Suppose A E jRnxn and
Let D = diag(d" ... , d
n
) be a nonsingular "scaling" matrix. Then
A
4l.
0 0
d,
0
)...
!b.
0
d,
D-'(X-' AX)D = D-' J D = j =
A
d
n
-
I
0
d
n
-
2
A-
d
n
d
n
-
I
0 0
)...
9.4. Geometric Aspects of the JCF 89
Appropriate choice of the di 's then yields any desired nonzero superdiagonal elements.
This result can also be interpreted in terms of the matrix X = [x\,..., x
n
] of eigenvectors
and principal vectors that reduces A to its JCF. Specifically, J is obtained from A via the
similarity transformation XD = \d\x\,..., d
n
x
n
}.
In a similar fashion, the reverse-order identity matrix (or exchange matrix)
Note that di mM( A A.,/ )
w
= ,-.
Definition 9.35. Let V be a vector space over F and suppose A : V > V is a linear
transformation. A subspace S c V is A-invariant if AS c S, where AS is defined as the
set {As : s e S}.
can be used to put the superdiagonal elements in the subdiagonal instead if that is desired:
The matrix X that reduces a matrix A e IR"
X
" (or C
nxn
) to a JCF provides a change of basis
with respect to which the matrix is diagonal or block diagonal. It is thus natural to expect an
associated direct sum decomposition of R. Such a decomposition is given in the following
theorem.
Theorem 9.34. Suppose A e R"
x
" has characteristic polynomial
and minimal polynomial
with A- i , . . . , A.
m
distinct. Then
9.4. Geometric Aspects of the JCF 89
Appropriate choice of the di's then yields any desired nonzero superdiagonal elements.
This result can also be interpreted in terms of the matrix X = [x[, ... ,x
n
] of eigenvectors
and principal vectors that reduces A to its lCF. Specifically, j is obtained from A via the
similarity transformation XD = [d[x[, ... , dnxn].
In a similar fashion, the reverse-order identity matrix (or exchange matrix)
0 0 I
0
p = pT = p-[ =
(9.18)
0 1
I 0 0
can be used to put the superdiagonal elements in the subdiagonal instead if that is desired:
A I 0 0 A 0 0
0 A 0 A 0
p-[
A
p=
0 1 A
0
A I A 0
0 0 A 0 0 A
The matrix X that reduces a matrix A E jH.nxn (or c
nxn
) to a lCF provides a change of basis
with respect to which the matrix is diagonal or block diagonal. It is thus natural to expect an
associated direct sum decomposition of jH.n. Such a decomposition is given in the following
theorem.
Theorem 9.34. Suppose A E jH.nxn has characteristic polynomial
n(A) = (A - A[)n) ... (A - Amt
m
and minimal polynomial
a(A) = (A - A[)V) '" (A - Am)V
m
with A I, ... , Am distinct. Then
jH.n = N(A - AlIt) E6 ... E6 N(A - AmItm
= N (A - A 1 I) v) E6 ... E6 N (A - Am I) Vm .
Note that dimN(A - AJ)Vi = ni.
Definition 9.35. Let V be a vector space over IF and suppose A : V --+ V is a linear
transformation. A subspace S ~ V is A -invariant if AS ~ S, where AS is defined as the
set {As: s E S}.
If V is taken to be R" over R, and S e R"
x
* is a matrix whose columns s\,..., s/t
span a /^-dimensional subspace <S, i.e., K(S) = <S, then <S is A-invariant if and only if there
exists M eR
kxk
such that
This follows easily by comparing the /th columns of each side of (9.19):
Example 9.36. The equation Ax = A* = xA defining a right eigenvector x of an eigenvalue
X says that * spans an A-invariant subspace (of dimension one).
Example 9.37. Suppose X block diagonalizes A, i.e.,
Rewriting in the form
we have that A A, = A", /,, / = 1, 2, so the columns of A, span an A-mvanant subspace.
Theorem 9.38. Suppose A e E"
x
".
7. Let p(A) = o/ + o?i A + + <x
q
A
q
be a polynomial in A. Then N(p(A)) and
7(p(A)) are A-invariant.
2. S isA-invariant if and only ifS
1
- is A
T
-invariant.
Theorem 9.39. If V isa vector space over F such that V = N\ 0 N
m
, where each
A// isA-invariant, then a basisfor V can be chosen with respect to which A hasa block
diagonal representation.
The Jordan canonical form is a special case of the above theorem. If A has distinct
eigenvalues A,,- as in Theorem 9.34, we could choose bases for N(A A.,-/)"' by SVD, for
example (note that the power n, could be replaced by v,). We would then get a block diagonal
representation for A with full blocks rather than the highly structured Jordan blocks. Other
such "canonical" forms are discussed in text that follows.
Suppose A" = [ X i , . . . , X
m
] e R"
n
xn
is such that X ÂX = diag(7i,. . . , J
m
), where
each Ji = diag(/,i,..., //*,.) and each /,* is a Jordan block corresponding to A, e A(A).
We could also use other block diagonal decompositions (e.g., via SVD), but we restrict our
attention here to only the Jordan block case. Note that A A", = A*,- /,, so by (9.19) the columns
of A", (i.e., the eigenvectors and principal vectors associated with A.,) span an A-invariant
subspace of W.
Finally, we return to the problem of developing a formula for e'
A
in the case that A
is not necessarily diagonalizable. Let 7, C"
x
"' be a Jordan basis for N(A
T
A.,/)"' .
Equivalently, partition
If V is taken to be ]Rn over Rand S E ]Rn xk is a matrix whose columns SI, ... , Sk
span a k-dimensional subspace S, i.e., R(S) = S, then S is A-invariant if and only if there
exists M E ]Rkxk such that
AS = SM. (9.19)
This follows easily by comparing the ith columns of each side of (9.19):
Example 9.36. The equation Ax = AX = x A defining a right eigenvector x of an eigenvalue
A says that x spans an A-invariant subspace (of dimension one).
Example 9.37. Suppose X block diagonalizes A, i.e.,
X-I AX = [ ~ J
2
].
Rewriting in the form
~ J,
we have that AX
i
= X;li, i = 1,2, so the columns of Xi span an A-invariant subspace.
Theorem 9.38. Suppose A E ]Rnxn.
1. Let peA) = CloI + ClIA + '" + ClqAq be a polynomial in A. Then N(p(A)) and
R(p(A)) are A-invariant.
2. S is A -invariant if and only if S 1. is A T -invariant.
Theorem 9.39. If V is a vector space over IF such that V = NI EB ... EB N
m
, where each
N; is A-invariant, then a basis for V can be chosen with respect to which A has a block
diagonal representation.
The Jordan canonical form is a special case of the above theorem. If A has distinct
eigenvalues Ai as in Theorem 9.34, we could choose bases for N(A - Ai/)n, by SVD, for
example (note that the power ni could be replaced by Vi). We would then get a block diagonal
representation for A with full blocks rather than the highly structured Jordan blocks. Other
such "canonical" forms are discussed in text that follows.
Suppose X = [Xl ..... Xm] E ] R ~ x n is such that X-I AX = diag(J1, ... , J
m
), where
each J
i
= diag(JiI,"" Jik,) and each Jik is a Jordan block corresponding to Ai E A(A).
We could also use other block diagonal decompositions (e.g., via SVD), but we restrict our
attention here to only the Jordan block case. Note that AXi = Xi J
i
, so by (9.19) the columns
of Xi (i.e., the eigenvectors and principal vectors associated with Ai) span an A-invariant
subspace of]Rn.
Finally, we return to the problem of developing a formula for e
l
A in the case that A
is not necessarily diagonalizable. Let Yi E <e
nxn
, be a Jordan basis for N (AT - A;lt.
Equivalently, partition
9.5. The Matrix Sign Function 91
compatibly. Then
In a similar fashion we can compute
which is a useful formula when used in conjunction with the result
for a k x k Jordan block 7, associated with an eigenvalue A. = A.,.
9.5 The Matrix Sign Function
In this section we give a very brief introduction to an interesting and useful matrix function
called the matrix sign function. It is a generalization of the sign (or signum) of a scalar. A
survey of the matrix sign function and some of its applications can be found in [15].
Definition 9.40. Let z E C with Re(z) ^ 0. Then the sign of z is defined by
Definition 9.41. Suppose A e C"
x
" has no eigenvalues on the imaginary axis, and let
be a Jordan canonical form for A, with N containing all Jordan blocks corresponding to the
eigenvalues of A in the left half-plane and P containing all Jordan blocks corresponding to
eigenvalues in the right half-plane. Then the sign of A, denoted sgn(A), is given by
9.S. The Matrix Sign Function
compatibly. Then
A = XJX-
I
= XJy
H
= [XI, ... , Xm] diag(JI, ... , J
m
) [Y
I
, , Ym]H
m
= LX;JiYi
H
.
i=1
In a similar fashion we can compute
m
etA = LXietJ;YiH,
i=1
which is a useful formula when used in conjunction with the result
A 0 0
eAt teAt
.lt
2
e
At
2!
0 A
0
eAt teAt
exp t
A 0
0 0
eAt
1
0 0 A
0 0
for a k x k Jordan block J
i
associated with an eigenvalue A = Ai.
9.5 The Matrix Sign Function
91
In this section we give a very brief introduction to an interesting and useful matrix function
called the matrix sign function. It is a generalization of the sign (or signum) of a scalar. A
survey of the matrix sign function and some of its applications can be found in [15].
Definition 9.40. Let z E C with Re(z) f= O. Then the sign of z is defined by
Re(z) {+1
sgn(z) = IRe(z) I = -1
ifRe(z) > 0,
ifRe(z) < O.
Definition 9.41. Suppose A E cnxn has no eigenvalues on the imaginary axis, and let
be a Jordan canonical form for A, with N containing all Jordan blocks corresponding to the
eigenvalues of A in the left half-plane and P containing all Jordan blocks corresponding to
eigenvalues in the right half-plane. Then the sign of A, denoted sgn(A), is given by
[
-/ 0] -I
sgn(A) = X 0 / X ,
where the negative and positive identity matrices are of the same dimensions as N and P,
respectively.
There are other equivalent definitions of the matrix sign function, but the one given
here is especially useful in deriving many of its key properties. The JCF definition of the
matrix sign function does not generally lend itself to reliable computation on a finite-word-
length digital computer. In fact, its reliable numerical calculation is an interesting topic in
its own right.
We state some of the more useful properties of the matrix sign function as theorems.
Their straightforward proofs are left to the exercises.
Theorem 9.42. Suppose A e C"
x
S = sgn(A). Then the following hold:
1. S is diagonalizable with eigenvalues equal to del.
2. S
2
= I.
3. AS = SA.
4. sgn(A") = (sgn(A))".
5. sgn(T-
l
AT) = T-
l
sgn(A)TforallnonsingularT e C"
x
".
6. sgn(cA) = sgn(c) sgn(A)/or all nonzero real scalars c.
Theorem 9.43. Suppose A e C"
x
S sgn(A). Then the following hold:
1. 7l(S /) is an A-invariant subspace corresponding to the left half-plane eigenvalues
of A (the negative invariant subspace).
2. R(S+/) is an A-invariant subspace corresponding to the right half-plane eigenvalues
of A (the positive invariant subspace).
3. negA = (/ S)/2 is a projection onto the negative invariant subspace of A.
4. posA = (/ + S)/2 is a projection onto the positive invariant subspace of A.
EXERCISES
1. Let A e C
nxn
have distinct eigenvalues AI, ..., X
n
with corresponding right eigen-
vectors Xi, ... ,x
n
and left eigenvectors y\, ..., y
n
, respectively. Let v e C" be an
arbitrary vector. Show that v can be expressed (uniquely) as a linear combination
of the right eigenvectors. Find the appropriate expression for v as a linear combination
of the left eigenvectors as well.
where the negative and positive identity matrices are of the same dimensions as N and p,
respectively.
There are other equivalent definitions of the matrix sign function, but the one given
here is especially useful in deriving many of its key properties. The JCF definition of the
matrix sign function does not generally lend itself to reliable computation on a finite-word-
length digital computer. In fact, its reliable numerical calculation is an interesting topic in
its own right.
We state some of the more useful properties of the matrix sign function as theorems.
Their straightforward proofs are left to the exercises.
Theorem 9.42. Suppose A E e
nxn
has no eigenvalues on the imaginary axis, and let
1. S is diagonalizable with eigenvalues equal to 1.
2. S2 = I.
3. AS = SA.
4. sgn(AH) = (sgn(AH.
5. sgn(T-1AT) = T-1sgn(A)T foralinonsingularT E e
nxn
.
6. sgn(cA) = sgn(c) sgn(A) for all nonzero real scalars c.
Theorem 9.43. Suppose A E e
nxn
has no eigenvalues on the imaginary axis, and let
I. R(S -l) is an A-invariant subspace corresponding to the left half-plane eigenvalues
of A (the negative invariant subspace).
2. R(S + l) is an A -invariant subspace corresponding to the right half-plane eigenvalues
of A (the positive invariant subspace).
3. negA == (l - S) /2 is a projection onto the negative invariant subspace of A.
4. posA == (l + S)/2 is a projection onto the positive invariant subspace of A.
EXERCISES
1. Let A E e
nxn
have distinct eigenvalues ),.1> , ),.n with corresponding right eigen-
vectors Xl, ... , Xn and left eigenvectors Yl, . , Yn, respectively. Let v E en be an
arbitrary vector. Show that v can be expressed (uniquely) as a linear combination
of the right eigenvectors. Find the appropriate expression for v as a linear combination
of the left eigenvectors as well.
Exercises 93
2. Suppose A C"
x
" is skew-Hermitian, i.e., A
H
= A. Prove that all eigenvalues of
a skew-Hermitian matrix must be pure imaginary.
3. Suppose A e C"
x
" is Hermitian. Let A be an eigenvalue of A with corresponding
right eigenvector x. Show that x is also a left eigenvector for A. Prove the same result
if A is skew-Hermitian.
5. Determine the eigenvalues, right eigenvectors and right principal vectors if necessary,
and (real) JCFs of the following matrices:
6. Determine the JCFs of the following matrices:
Find a nonsingular matrix X such that X
1
AX = J, where J is the JCF
Hint: Use[ 1 1 l]
r
as an eigenvector. The vectors [0 1 l]
r
and[ l 0 0]
r
are both eigenvectors, but then the equation (A /)jc
(2)
= x
(1)
can't be solved.
8. Show that all right eigenvectors of the Jordan block matrix in Theorem 9.30 must be
multiples of e\ e R*. Characterize all left eigenvectors.
9. Let A e R"
x
" be of the form A = xy
T
, where x, y e R" are nonzero vectors with
x
T
y = 0. Determine the JCF of A.
10. Let A e R"
xn
be of the form A = / + xy
T
, where x, y e R" are nonzero vectors
with x
T
y = 0. Determine the JCF of A.
11. Suppose a matrix A e R
16x 16
has 16 eigenvalues at 0 and its JCF consists of a single
Jordan block of the form specified in Theorem 9.22. Suppose the small number 10~
16
is added to the (16,1) element of J. What are the eigenvalues of this slightly perturbed
matrix?
4. Suppose a matrix A R
5x5
has eigenvalues {2, 2, 2, 2, 3}. Determine all possible
JCFs for A.
7. Let
Exercises 93
2. Suppose A E rc
nxn
is skew-Hermitian, i.e., AH = -A. Prove that all eigenvalues of
a skew-Hermitian matrix must be pure imaginary.
3. Suppose A E rc
nxn
is Hermitian. Let A be an eigenvalue of A with corresponding
right eigenvector x. Show that x is also a left eigenvector for A. Prove the same result
if A is skew-Hermitian.
4. Suppose a matrix A E lR.
5x5
has eigenvalues {2, 2, 2, 2, 3}. Determine all possible
JCFs for A.
5. Determine the eigenvalues, right eigenvectors and right principal vectors if necessary,
and (real) JCFs of the following matrices:
[
2 -1 ]
(a) 1 0 '
6. Determine the JCFs of the following matrices:
<a) U j n
-2
-1
2
=n
7. Let
A = [H -1]
2 2"
Find a nonsingular matrix X such that X-I AX = J, where J is the JCF
J = [ ~ ~ ~ ] .
001
Hint: Use[-1 1 - I]T as an eigenvector. The vectors [0 -If and[1 0 of
are both eigenvectors, but then the equation (A - I)x(2) = x(1) can't be solved.
8. Show that all right eigenvectors of the Jordan block matrix in Theorem 9.30 must be
multiples of el E lR.
k
. Characterize all left eigenvectors.
9. Let A E lR.
nxn
be of the form A = xyT, where x, y E lR.
n
are nonzero vectors with
x
T
y = O. Determine the JCF of A.
10. Let A E lR.
nxn
be of the form A = 1+ xyT, where x, y E lR.
n
are nonzero vectors
with x
T
y = O. Determine the JCF of A.
11. Suppose a matrix A E lR.
16x
16 has 16 eigenvalues at 0 and its JCF consists of a single
Jordan block of the form specified in Theorem 9.22. Suppose the small number 10-
16
is added to the (16,1) element of J. What are the eigenvalues of this slightly perturbed
matrix?
12. Show that every matrix A e R"
x
" can be factored in the form A = Si$2, where Si
and 2 are real symmetric matrices and one of them, say Si, is nonsingular.
Hint: Suppose A = X J X ~
l
is a reduction of A to JCF and suppose we can construct
the "symmetric factorization" of J . Then A = ( X S i X
T
) ( X ~
T
S
2
X ~
l
) would be the
required symmetric factorization of A. Thus, it suffices to prove the result for the
JCF. The transformation P in (9.18) is useful.
13. Prove that every matrix A e W
x
" is similar to its transpose and determine a similarity
transformation explicitly.
Hint: Use the factorization in the previous exercise.
14. Consider the block upper triangular matrix
where A e M"
xn
and A
n
e R
kxk
with 1 < k < n. Suppose A
u
^ 0 and that we
want to block diagonalize A via the similarity transformation
where X e R*
x
< - *), i.e.,
Find a matrix equation that X must satisfy for this to be possible. If n = 2 and k = 1,
what can you say further, in terms of AU and A 22, about when the equation for X is
solvable?
17. Suppose A e C"
xn
has all its eigenvalues in the left half- plane. Prove that
sgn(A) = - /.
12. Show that every matrix A E jRnxn can be factored in the form A = SIS2, where SI
and S2 are real symmetric matrices and one of them, say S1, is nonsingular.
Hint: Suppose A = Xl X-I is a reduction of A to JCF and suppose we can construct
the "symmetric factorization" of 1. Then A = (X SIXT)(X-
T
S2X-I) would be the
required symmetric factorization of A. Thus, it suffices to prove the result for the
JCF. The transformation P in (9.18) is useful.
13. Prove that every matrix A E jRn xn is similar to its transpose and determine a similarity
transformation explicitly.
Hint: Use the factorization in the previous exercise.
14. Consider the block upper triangular matrix
A _ [ All
- 0
Al2 ]
A22 '
where A E jRnxn and All E jRkxk with 1 ::s: k ::s: n. Suppose Al2 =1= 0 and that we
want to block diagonalize A via the similarity transformation
where X E IRkx(n-k), i.e.,
T-IAT = [A 011 0 ]
A22 .
Find a matrix equation that X must satisfy for this to be possible. If n = 2 and k = 1,
what can you say further, in terms of All and A22, about when the equation for X is
solvable?
17. Suppose A E en xn has all its eigenvalues in the left half-plane. Prove that
sgn(A) = -1.
Chapter 10
Canonical Forms
10.1 Some Basic Canonical Forms
Problem: Let V and W be vector spaces and suppose A : V > W is a linear transformation.
Find bases in V and W with respect to which Mat A has a "simple form" or "canonical
form." In matrix terms, if A e R
mxn
, find P e R
xm
and Q e R
n
n
xn
such that PAQ has a
"canonical form." The transformation A M PAQ is called an equivalence; it is called an
orthogonal equivalence if P and Q are orthogonal matrices.
Remark 10.1. We can also consider the case A e C
m xn
and unitary equivalence if P and
< 2 are unitary.
Two special cases are of interest:
1. If W = V and < 2 = P"
1
, the transformation A H> PAP"
1
is called a similarity.
2 . If W = V and if Q = P
T
is orthogonal, the transformation A i- PAP
T
is called
an orthogonal similarity (or unitary similarity in the complex case).
The following results are typical of what can be achieved under a unitary similarity. If
A = A
H
6 C"
x
" has eigenvalues AI, . . . , A
n
, then there exists a unitary matrix 7 such that
U
H
AU D, where D = di ag( A. j , . . . , A.
n
). This is proved in Theorem 10.2 . What other
matrices are "diagonalizable" under unitary similarity? The answer is given in Theorem
10.9, where it is proved that a general matrix A e C"
x
" is unitarily similar to a diagonal
matrix if and only if it is normal (i.e., AA
H
= A
H
A). Normal matrices include Hermitian,
skew-Hermitian, and unitary matrices (and their "real" counterparts: symmetric, skew-
symmetric, and orthogonal, respectively), as well as other matrices that merely satisfy the
definition, such as A = [ _
a
b
^1 for real scalars a and b. If a matrix A is not normal, the
most "diagonal" we can get is the JCF described in Chapter 9.
Theorem 10.2. Let A = A
H
e C"
x
" have (real) eigenvalues A. I, . . . , X
n
. Then there
exists a unitary matrix X such that X
H
AX = D = diag(A.j , . . . , X
n
) (the columns ofX are
orthonormal eigenvectors for A).
95
Chapter 10
Canonical Forms
10.1 Some Basic Canonical Forms
Problem: Let V and W be vector spaces and suppose A : V ---+ W is a linear transformation.
Find bases in V and W with respect to which Mat A has a "simple form" or "canonical
form." In matrix terms, if A E IR
mxn
, find P E lR;;:xm and Q E l R ~ x n such that P AQ has a
"canonical form." The transformation A f--+ P AQ is called an equivalence; it is called an
orthogonal equivalence if P and Q are orthogonal matrices.
Remark 10.1. We can also consider the case A E e
mxn
and unitary equivalence if P and
Q are unitary.
Two special cases are of interest:
1. If W = V and Q = p-
1
, the transformation A f--+ PAP-I is called a similarity.
2. If W = V and if Q = pT is orthogonal, the transformation A f--+ P ApT is called
an orthogonal similarity (or unitary similarity in the complex case).
The following results are typical of what can be achieved under a unitary similarity. If
A = A H E en xn has eigenvalues AI, ... , An, then there exists a unitary matrix U such that
U
H
AU = D, where D = diag(AJ, ... , An). This is proved in Theorem 10.2. What other
matrices are "diagonalizable" under unitary similarity? The answer is given in Theorem
10.9, where it is proved that a general matrix A E e
nxn
is unitarily similar to a diagonal
matrix if and only if it is normal (i.e., AA H = AHA). Normal matrices include Hermitian,
skew-Hermitian, and unitary matrices (and their "real" counterparts: symmetric, skew-
symmetric, and orthogonal, respectively), as well as other matrices that merely satisfy the
definition, such as A = _ ~ !] for real scalars a and h. If a matrix A is not normal, the
most "diagonal" we can get is the JCF described in Chapter 9.
Theorem 10.2. Let A = A H E en xn have (real) eigenvalues AI, ... ,An. Then there
exists a unitary matrix X such that X
H
AX = D = diag(Al, ... , An) (the columns of X are
orthonormal eigenvectors for A).
95
96 Chapter 10. Canonical Forms
Proof: Let x\ be a right eigenvector corresponding to X\, and normalize it such that xf*x\ =
1. Then there exist n 1 additional vectors x
2
, ..., x
n
such that X = [x\,..., x
n
] =
[x\ X
2
] is unitary. Now
Then xÛ
2
= 0 (/ k) means that x
f
is orthogonal to each of the n k columns of U
2
.
But the latter are orthonormal since they are the last n k rows of the unitary matrix U.
Thus, [Xi f/2] is unitary. D
The construction called for in Theorem 10.2 is then a special case of Theorem 10.3
for k = 1. We illustrate the construction of the necessary Householder matrix for k 1.
For simplicity, we consider the real case. Let the unit vector x\ be denoted by [i, ..., %
n
]
T
.
In (10.1) we have used the fact that Ax\ = k\x\. When combined with the fact that
x"xi = 1, we get A-i remaining in the (l,l)-block. We also get 0 in the (2,l)-block by
noting that x\ is orthogonal to all vectors in X
2
. In (10.2), we get 0 in the (l,2)-block by
noting that X
H
AX is Hermitian. The proof is completed easily by induction upon noting
that the (2,2)-block must have eigenvalues X
2
,..., A.
n
. D
Given a unit vector x\ e E", the construction of X
2
e ]R"
X
("-
1
) such that X
[x\ X
2
] is orthogonal is frequently required. The construction can actually be performed
quite easily by means of Householder (or Givens) transformations as in the proof of the
following general result.
Theorem 10.3. Let X\ e C
nxk
have orthonormal columns and suppose U is a unitary
matrix such that UX\ = \
0
1, where R C
kxk
is upper triangular. Write U
H
= [U\ U
2
]
with Ui C
nxk
. Then [Xi U
2
] is unitary.
Proof: Let X\ = [x\,..., Xk]. Construct a sequence of Householder matrices (also known
as elementary reflectors) H\,..., H
k
in the usual way (see below) such that
where R is upper triangular (and nonsingular since x\, ..., Xk are orthonormal). Let U =
H
k
...H
v
. Then U
H
= / / ,- H
k
and
Proof' Let XI be a right eigenvector corresponding to AI, and normalize it such that XI =
1. Then there exist n - 1 additional vectors X2, ... , Xn such that X = (XI, ... , xn] =
[XI X
2
] is unitary. Now
XHAX = [
xH
] A [XI X2] = [

]
I
XH
XfAxl XfAX
2
2
=[
Al
]
(10.1)
0 XfAX
2
=[
Al 0
l
(10.2)
0
XfAX
z
In (l0.1) we have used the fact that AXI = AIXI. When combined with the fact that
XI = 1, we get Al remaining in the (l,I)-block. We also get 0 in the (2, I)-block by
noting that XI is orthogonal to all vectors in Xz. In (10.2), we get 0 in the (l,2)-block by
noting that XH AX is Hermitian. The proof is completed easily by induction upon noting
that the (2,2)-block must have eigenvalues A2, ... , An. 0
Given a unit vector XI E JRn, the construction of X
z
E JRnx(n-l) such that X =
[XI X
2
] is orthogonal is frequently required. The construction can actually be performed
quite easily by means of Householder (or Givens) transformations as in the proof of the
following general result.
Theorem 10.3. Let XI E C
nxk
have orthonormal columns and suppose V is a unitary
matrix such that V X I = [ where R E C
kxk
is upper triangular. Write V H = [VI Vz]
with VI E C
nxk
. Then [XI V
2
] is unitary.
Proof: Let X I = [XI, ... ,xd. Construct a sequence of Householder matrices (also known
as elementary reflectors) HI, ... , Hk in the usual way (see below) such that
Hk ... HdxI, ... , xd = [ l
where R is upper triangular (and nonsingular since XI, ... , Xk are orthonormal). Let V =
Hk'" HI. Then VH = HI'" Hk and
Then X
i
H
U2 = 0 (i E means that Xi is orthogonal to each of the n - k columns of V2.
But the latter are orthonormal since they are the last n - k rows of the unitary matrix U.
Thus. [XI U2] is unitary. 0
The construction called for in Theorem 10.2 is then a special case of Theorem 10.3
for k = 1. We illustrate the construction of the necessary Householder matrix for k = 1.
For simplicity, we consider the real case. Let the unit vector XI be denoted by .. , ,
10.1. Some Basic Canonical Forms 97
Then the necessary Householder matrix needed for the construction of X^ is given by
U = I 2uu
+
= I -ÛU
T
, where u = [t-\ 1, 2, ]
r
- It can easily be checked
that U is symmetric and U
T
U = U
2
= I, so U is orthogonal. To see that U effects the
necessary compression of j ci, it is easily verified that U
T
U = 2 2i and U
T
X\ = 1 1.
Thus,
Further details on Householder matrices, including the choice of sign and the complex case,
can be consulted in standard numerical linear algebra texts such as [7], [11], [23], [25].
The real version of Theorem 10.2 is worth stating separately since it is applied fre-
quently in applications.
T
e E
nxn
have eigenvalues k\, ... ,X
n
. Then there exists an
orthogonal matrix X e W
lxn
(whose columns are orthonormal eigenvectors of A) such that
X
T
AX = D = diag(Xi, . . . , X
n
).
Note that Theorem 10.4 implies that a symmetric matrix A (with the obvious analogue
from Theorem 10.2 for Hermitian matrices) can be written
which is often called the spectral representation of A. In fact, A in (10.3) is actually a
weighted sum of orthogonal projections P, (onto the one-dimensional eigenspaces corre-
sponding to the A., 's), i.e.,
where P, = PUM
x
i
x
f =
x
i
x
j since xj xi 1.
The following pair of theorems form the theoretical foundation of the double-Francis-
QR algorithm used to compute matrix eigenvalues in a numerically stable and reliable way.
10.1. Some Basic Canonical Forms 97
Then the necessary Householder matrix needed for the construction of X
2
is given by
U = I - 2uu+ = I - +uu
T
, where u = [';1 1, ';2, ... , ';nf. It can easily be checked
u u
that U is symmetric and U
T
U = U
2
= I, so U is orthogonal. To see that U effects the
necessary compression of Xl, it is easily verified that u
T
u = 2 2';1 and u
T
Xl = 1 ';1.
Thus,
Further details on Householder matrices, including the choice of sign and the complex case,
can be consulted in standard numerical linear algebra texts such as [7], [11], [23], [25].
The real version of Theorem 10.2 is worth stating separately since it is applied fre-
quently in applications.
Theorem 10.4. Let A = AT E jRnxn have eigenvalues AI, ... , An. Then there exists an
orthogonal matrix X E jRn xn (whose columns are orthonormal eigenvectors of A) such that
XT AX = D = diag(Al, ... , An).
Note that Theorem 10.4 implies that a symmetric matrix A (with the obvious analogue
from Theorem 10.2 for Hermitian matrices) can be written
n
A = XDX
T
= LAiXiXT,
(10.3)
i=1
which is often called the spectral representation of A. In fact, A in (10.3) is actually a
weighted sum of orthogonal projections Pi (onto the one-dimensional eigenspaces corre-
sponding to the Ai'S), i.e.,
n
A = LAiPi,
i=l
where Pi = PR(x;) = xiXt = xixT since xT Xi = 1.
The following pair of theorems form the theoretical foundation of the double-Francis-
QR algorithm used to compute matrix eigenvalues in a numerically stable and reliable way.
Theorem 10.5 (Schur). Let A e C"
x
". Then there exists a unitary matrix U such that
U
H
AU = T, where T is upper triangular.
Proof: The proof of this theorem is essentially the same as that of Theorem 10.2 except that
in this case (using the notation U rather than X) the (l,2)-block wf AU2 is not 0. D
In the case of A e R"
x
", it is thus unitarily similar to an upper triangular matrix, but
if A has a complex conjugate pair of eigenvalues, then complex arithmetic is clearly needed
to place such eigenvalues on the diagonal of T. However, the next theorem shows that every
A e W
xn
is also orthogonally similar (i.e., real arithmetic) to a quasi-upper-triangular
matrix. A quasi-upper-triangular matrix is block upper triangular with 1 x 1 diagonal
blocks corresponding to its real eigenvalues and 2x2 diagonal blocks corresponding to its
complex conjugate pairs of eigenvalues.
Theorem 10.6 (Murnaghan-Wintner). Let A e R"
x
". Then there exists an orthogonal
matrix U such that U
T
AU = S, where S is quasi-upper-triangular.
Definition 10.7. The triangular matrix T in Theorem 10.5 is called a Schur canonical
form or Schur form. The quasi-upper-triangular matrix S in Theorem 10.6 is called a real
Schur canonical form or real Schur form (RSF). The columns of a unitary [orthogonal]
matrix U that reduces a matrix to [real] Schur form are called Schur vectors.
Example 10.8. The matrix
is in RSF. Its real JCF is
Note that only the first Schur vector (and then only if the corresponding first eigenvalue
is real if U is orthogonal) is an eigenvector. However, what is true, and sufficient for virtually
all applications (see, for example, [17]), is that the first k Schur vectors span the same A-
invariant subspace as the eigenvectors corresponding to the first k eigenvalues along the
diagonal of T (or S).
While every matrix can be reduced to Schur form (or RSF), it is of interest to know
when we can go further and reduce a matrix via unitary similarity to diagonal form. The
following theorem answers this question.
Theorem 10.9. A matrix A e C"
x
" is unitarily similar to a diagonal matrix if and only if
A is normal (i.e., A
H
A = AA
H
).
Proof: Suppose U is a unitary matrix such that U
H
AU = D, where D is diagonal. Then
so A is normal.
Theorem 10.5 (Schur). Let A E c
nxn
. Then there exists a unitary matrix U such that
U
H
AU = T, where T is upper triangular.
Proof: The proof of this theorem is essentially the same as that of Theorem lO.2 except that
in this case (using the notation U rather than X) the (l,2)-block ur AU
2
is not O. 0
In the case of A E IR
n
xn , it is thus unitarily similar to an upper triangular matrix, but
if A has a complex conjugate pair of eigenvalues, then complex arithmetic is clearly needed
to place such eigenValues on the diagonal of T. However, the next theorem shows that every
A E IR
nxn
is also orthogonally similar (i.e., real arithmetic) to a quasi-upper-triangular
matrix. A quasi-upper-triangular matrix is block upper triangular with 1 x 1 diagonal
blocks corresponding to its real eigenvalues and 2 x 2 diagonal blocks corresponding to its
complex conjugate pairs of eigenvalues.
Theorem 10.6 (Murnaghan-Wintner). Let A E IR
n
xn. Then there exists an orthogonal
T
AU = S, where S is quasi-upper-triangular.
Definition 10.7. The triangular matrix T in Theorem 10.5 is called a Schur canonical
form or Schur fonn. The quasi-upper-triangular matrix S in Theorem 10.6 is called a real
Schur canonical form or real Schur fonn (RSF). The columns of a unitary [orthogonal}
matrix U that reduces a matrix to [real} Schur fonn are called Schur vectors.
Example 10.8. The matrix
s ~ [
-2 5
n
-2 4
0 0
is in RSF. Its real JCF is
h[
1
n
-1 1
0 0
Note that only the first Schur vector (and then only if the corresponding first eigenvalue
is real if U is orthogonal) is an eigenvector. However, what is true, and sufficient for virtually
all applications (see, for example, [17]), is that the first k Schur vectors span the same A-
invariant subspace as the eigenvectors corresponding to the first k eigenvalues along the
diagonal of T (or S).
While every matrix can be reduced to Schur form (or RSF), it is of interest to know
when we can go further and reduce a matrix via unitary similarity to diagonal form. The
following theorem answers this question.
Theorem 10.9. A matrix A E c
nxn
is unitarily similar to a diagonal matrix if and only if
A is normal (i.e., AH A = AA
H
).
Proof: Suppose U is a unitary matrix such that U
H
AU = D, where D is diagonal. Then
AAH = U VUHU VHU
H
= U DDHU
H
== U DH DU
H
== AH A
so A is normal.
10.2. Definite Matrices 99
Conversely, suppose A is normal and let U be a unitary matrix such that U
H
AU = T,
where T is an upper triangular matrix (Theorem 10.5). Then
It is then a routine exercise to show that T must, in fact, be diagonal. D
10.2 Definite Matrices
Definition 10.10. A symmetric matrix A e W
xn
is
1. positive definite if and only ifx
T
Ax > Qfor all nonzero x G W
1
. We write A > 0.
2. nonnegative definite (or positive semidefinite) if and only if X
T
Ax > 0 for all
nonzero x e W. We write A > 0.
3. negative definite ifA is positive definite. We write A < 0.
4. nonpositive definite (or negative semidefinite) if A is nonnegative definite. We
write A < 0.
Also, if A and B are symmetric matrices, we write A > B if and only if A B > 0 or
B A < 0. Similarly, we write A > B if and only ifA B>QorB A < 0.
Remark 10.11. If A e C"
x
" is Hermitian, all the above definitions hold except that
superscript //s replace Ts. Indeed, this is generally true for all results in the remainder of
this section that may be stated in the real case for simplicity.
Remark 10.12. If a matrix is neither definite nor semidefinite, it is said to be indefinite.
H
e C
nxn
with eigenvalues X
{
> A
2
> > A
n
. Then for all
x eC",
Proof: Let U be a unitary matrix that diagonalizes A as in Theorem 10.2. Furthermore,
let v = U
H
x, where x is an arbitrary vector in C
M
, and denote the components of y by
j]i, i n. Then
But clearly
Conversely, suppose A is normal and let U be a unitary matrix such that U H A U = T,
where T is an upper triangular matrix (Theorem 10.5). Then
It is then a routine exercise to show that T must, in fact, be diagonal. 0
10.2 Definite Matrices
Definition 10.10. A symmetric matrix A E lR.
nxn
is
1. positive definite if and only if x T Ax > 0 for all nonzero x E lR.
n
. We write A > O.
2. nonnegative definite (or positive semidefinite) if and only if x
T
Ax :::: 0 for all
nonzero x E lR.
n
We write A :::: O.
3. negative definite if - A is positive definite. We write A < O.
4. nonpositive definite (or negative semidefinite) if -A is nonnegative definite. We
write A ~ O.
Also, if A and B are symmetric matrices, we write A > B if and only if A - B > 0 or
B - A < O. Similarly, we write A :::: B if and only if A - B :::: 0 or B - A ~ O.
Remark 10.11. If A E e
nxn
is Hermitian, all the above definitions hold except that
superscript H s replace T s. Indeed, this is generally true for all results in the remainder of
this section that may be stated in the real case for simplicity.
Remark 10.12. If a matrix is neither definite nor semidefinite, it is said to be indefinite.
Theorem 10.13. Let A = AH E e
nxn
with eigenvalues AI :::: A2 :::: ... :::: An. Thenfor all
x E en,
Proof: Let U be a unitary matrix that diagonalizes A as in Theorem 10.2. Furthermore,
let y = U H x, where x is an arbitrary vector in en, and denote the components of y by
11;, i En. Then
But clearly
n
x
H
Ax = (U
H
X)H U
H
AU(U
H
x) = yH Dy = LA; 111;12.
n
LA; 11'/;12 ~ AlyH Y = AIX
H
X
;=1
;=1
and
from which the theorem follows. D
Remark 10.14. The ratio ^^ for A = A
H
< = C
nxn
and nonzero jc e C" is called the
Rayleigh quotient of jc. Theorem 10.13 provides upper (AO and lower (A.
w
) bounds for
the Rayleigh quotient. If A = A
H
e C"
x
" is positive definite, X
H
Ax > 0 for all nonzero
x E C", soO < X
n
< < A. I.
Corollary 10.15. Let A e C"
x
". Then \\A\\
2
= ^
m
(A
H
A}.
Proof: For all x C" we have
Let jc be an eigenvector corresponding to X
max
(A
H
A). Then ^pjp
2
= ^^(A" A) , whence
Definition 10.16. A principal submatrix of an nxn matrix A is the (n k)x(n k) matrix
that remains by deleting k rows and the corresponding k columns. A leading principal
submatrix of order n k is obtained by deleting the last k rows and columns.
Theorem 10.17. A symmetric matrix A e E"
x
" is positive definite if and only if any of the
following three equivalent conditions hold:
1. The determinants of all leading principal submatrices of A are positive.
2. All eigenvalues of A are positive.
3. A can be written in the formM
T
M, where M e R"
x
" is nonsingular.
Theorem 10.18. A symmetric matrix A R"
x
" is nonnegative definite if and only if any
of the following three equivalent conditions hold:
1. The determinants of all principal submatrices of A are nonnegative.
2. All eigenvalues of A are nonnegative.
3. A can be written in the formM
T
M, where M 6 R
ix
" and k > rank(A) rank(M) .
Remark 10.19. Note that the determinants of all principal eubmatrioes muet bQ nonnogativo
in Theorem 10.18.1, not just those of the leading principal submatrices. For example,
consider the matrix A [
0
_
l
1. The determinant of the 1x1 leading submatrix is 0 and
the determinant of the 2x2 leading submatrix is also 0 (cf. Theorem 10.17). However, the
and
n
LAillJilZ::: AnyHy = An
xHx
,
i=l
from which the theorem follows. 0
Remark 10.14. The ratio XHHAx for A = AH E e
nxn
and nonzero x E en is called the
x x
Rayleigh quotient of x. Theorem 1O.l3 provides upper (A 1) and lower (An) bounds for
the Rayleigh quotient. If A = AH E e
nxn
is positive definite, x
H
Ax > 0 for all nonzero
x E en, so 0 < An ::::: ... ::::: A I.
I
Corollary 10.15. Let A E e
nxn
. Then IIAII2 = Ar1ax(AH A).
Proof: For all x E en we have
I
Let x be an eigenvector corresponding to Amax (A H A). Then = Ar1ax (A H A), whence
IIAxll2 ! H
IIAliz = max --= Amax{A A). 0
xfO IIxll2
Definition 10.16. A principal submatrixofan n x n matrix A is the (n -k) x (n -k) matrix
that remains by deleting k rows and the corresponding k columns. A leading principal
submatrix of order n - k is obtained by deleting the last k rows and columns.
Theorem 10.17. A symmetric matrix A E is positive definite if and only if any of the
following three equivalent conditions hold:
1. The determinants of all leading principal submatrices of A are positive.
2. All eigenvalues of A are positive.
3. A can be written in the form MT M, where M E xn is nonsingular.
Theorem 10.18. A symmetric matrix A E xn is nonnegative definite if and only if any
of the following three equivalent conditions hold:
1. The determinants of all principal submatrices of A are nonnegative.
2. All eigenvalues of A are nonnegaTive.
3. A can be wrirren in [he/orm MT M, where M E IRb<n and k ranlc(A) "" ranlc(M).
R.@mllrk 10.19. Not@th!ltthl!dl!termin!lntl:ofnllprincip!ll "ubm!ltriC[!!l mu"t bB nonnBgmivB
in Theorem 10.18.1, not just those of the leading principal submatrices. For example,
consider the matrix A = _ The determinant of the I x 1 leading submatrix is 0 and
the determinant of the 2 x 2 leading submatrix is also 0 (cf. Theorem 10.17). However, the
principal submatrix consisting of the (2,2) element is, in fact, negative and A is nonpositive
definite.
Remark 10.20. The factor M in Theorem 10.18.3 is not unique. For example, if
Recall that A > B if the matrix A B is nonnegative definite. The following
theorem is useful in "comparing" symmetric matrices. Its proof is straightforward from
basic definitions.
Theorem 10.21. Let A, B e R
nxn
be symmetric.
1. If A >BandMe R
nxm
, then M
T
AM > M
T
BM.
2. If A >B and M e R
nxm
, then M
T
AM > M.
T
BM.
j m
The following standard theorem is stated without proof (see, for example, [16, p.
181]). It concerns the notion of the "square root" of a matrix. That is, if A E"
xn
, we say
that S e R
nx
" is a square root of A if S
2
A. In general, matrices (both symmetric and
nonsymmetric) have infinitely many square roots. For example, if A = /2, any matrix S of
the form [
c
s
*
e
e
_
c
s
9
e
] is a square root.
Theorem 10.22. Let A e R"
x
" be nonnegative definite. Then A has a unique nonnegative
definite square root S. Moreover, SA = AS and rankS = rank A (and hence S is positive
definite if A is positive definite).
A stronger form of the third characterization in Theorem 10.17 is available and is
known as the Cholesky factorization. It is stated and proved below for the more general
Hermitian case.
Theorem 10.23. Let A e <C
nxn
be Hermitian and positive definite. Then there exists a
unique nonsingular lower triangular matrix L with positive diagonal elements such that
A = LL
H
.
Proof: The proof is by induction. The case n = 1 is trivially true. Write the matrix A in
the form
By our induction hypothesis, assume the result is true for matrices of order n 1 so that B
may be written as B = L\L^, where L\ e C
1
-""
1
^""^ is nonsingular and lower triangular
then M can be
principal submatrix consisting of the (2,2) element is, in fact, negative and A is nonpositive
definite.
Remark 10.20. The factor M in Theorem 10.18.3 is not unique. For example, if
then M can be
[1 0], [ fz
-ti
o [ ~ 0]
o l ~ 0 , ...
v'3 0
Recall that A :::: B if the matrix A - B is nonnegative definite. The following
theorem is useful in "comparing" symmetric matrices. Its proof is straightforward from
basic definitions.
Theorem 10.21. Let A, B E jRnxn be symmetric.
1. 1f A :::: Band M E jRnxm, then MT AM :::: MT BM.
2. If A> Band M E j R ~ x m , then MT AM> MT BM.
The following standard theorem is stated without proof (see, for example, [16, p.
181]). It concerns the notion of the "square root" of a matrix. That is, if A E lR.
nxn
, we say
that S E jRn xn is a square root of A if S2 = A. In general, matrices (both symmetric and
nonsymmetric) have infinitely many square roots. For example, if A = lz, any matrix S of
h
" [COSO Sino] .
t e 10rm sinO _ cosO IS a square root.
Theorem 10.22. Let A E lR.
nxn
be nonnegative definite. Then A has a unique nonnegative
definite square root S. Moreover, SA = AS and rankS = rankA (and hence S is positive
definite if A is positive definite).
A stronger form of the third characterization in Theorem 10.17 is available and is
known as the Cholesky factorization. It is stated and proved below for the more general
Hermitian case.
nxn
be Hermitian and positive definite. Then there exists a
unique nonsingular lower triangular matrix L with positive diagonal elements such that
A = LLH.
Proof: The proof is by induction. The case n = 1 is trivially true. Write the matrix A in
the form
By our induction hypothesis, assume the result is true for matrices of order n - 1 so that B
may be written as B = L1Lf, where Ll E c(n-l)x(n-l) is nonsingular and lower triangular
102 Chapt er 10. Ca n o n i c a l Forms
with positive diagonal elements. It remains to prove that we can write the n x n matrix A
in the form
Theorem 10.24. Let A C*
7 1
. Then there exist matrices P e C
xm
and Q e C"
n
x
" such
that
Proof: A classical proof can be consulted in, for example, [21, p. 131]. Alternatively,
suppose A has an SVD of the form (5.2) in its complex version. Then
Note that the greater freedom afforded by the equivalence transformation of Theorem
10.24, as opposed to the more restrictive situation of a similarity transformation, yields a
far "simpler" canonical form (10.4). However, numerical procedures for computing such
an equivalence directly via, say, Gaussian or elementary row and column operations, are
generally unreliable. The numerically preferred equivalence is, of course, the unitary equiv-
alence known as the SVD. However, the SVD is relatively expensive to compute and other
canonical forms exist that are intermediate between (10.4) and the SVD; see, for example
[7, Ch. 5], [4, Ch. 2]. Two such forms are stated here. They are more stably computable
than (10.4) and more efficiently computable than a ful l SVD. Many similar results are also
available.
where a is positive. Performing the indicated matrix multiplication and equating the cor-
responding submatrices, we see that we must have L\c = b and a
nn
= C
H
C + a
2
. Clearly
c is given simply by c = L^b. Substituting in the expression involving a, we find
a
2
= a
nn
b
H
L\
H
L\
l
b = a
nn
b
H
B~
l
b (= the Schur complement of B in A). But we
know that
Since det (fi ) > 0, we must have a
nn
b
H
B
l
b > 0. Choosing a to be the positive square
root of b
H
B~
l
b completes the proof. D
with positive diagonal elements. It remains to prove that we can write the n x n matrix A
in the form
b ] = [L
J
0 ] [Lf c J,
ann c a 0 a
where a is positive. Performing the indicated matrix multiplication and equating the cor-
responding submatrices, we see that we must have L I C = b and ann = c
H
c + a
2
Clearly
c is given simply by c = C,lb. Substituting in the expression involving a, we find
a
2
= ann - b
H
LIH L11b = ann - b
H
B-1b (= the Schur complement of B in A). But we
know that
o < det(A) = det [
b ] = det(B) det(a
nn
_ b
H
B-1b).
ann
Since det(B) > 0, we must have ann - b
H
B-1b > O. Choosing a to be the positive square
root of ann - b
H
B-1b completes the proof. 0
Theorem 10.24. Let A E c;,xn. Then there exist matrices P E C:
xm
and Q E such
that

(l0.4)
Proof: A classical proof can be consulted in, for example, [21, p. 131]. Alternatively,
suppose A has an SVD of the form (5.2) in its complex version. Then
[
S-l 0 ] [ U
H
] [I 0 ]
o I Uf AV = 0 0 .
Take P = [ 'f [I ] and Q = V to complete the proof. 0
Note that the greater freedom afforded by the equivalence transformation of Theorem
10.24, as opposed to the more restrictive situation of a similarity transformation, yields a
far "simpler" canonical form (10.4). However, numerical procedures for computing such
an equivalence directly via, say, Gaussian or elementary row and column operations, are
generally unreliable. The numerically preferred equivalence is, of course, the unitary equiv-
alence known as the SVD. However, the SVD is relatively expensive to compute and other
canonical forms exist that are intermediate between (l0.4) and the SVD; see, for example
[7, Ch. 5], [4, Ch. 2]. Two such forms are stated here. They are more stably computable
than (lOA) and more efficiently computable than a full SVD. Many similar results are also
available.
10.3. Equivalence Transformations and Congruence 103
Theorem 10.25 (Complete Orthogonal Decomposition). Let A e C
x
". Then there exist
unitary matrices U e C
mxm
and V e C
nxn
such that
where R e ,
r
r
xr
is upper (or lower) triangular with positive diagonal elements.
Theorem 10.26. Let A e C
x
". Then there exists a unitary matrix Q e C
mxm
and a
permutation matrix Fl e C"
x
" such that
where R E C
r
r
xr
is upper triangular and S e C
r x(
"
r)
is arbitrary but in general nonzero.
Remark 10.27. When A has full column rank but is "near" a rank deficient matrix,
various rank revealing QR decompositions are available that can sometimes detect such
phenomena at a cost considerably less than a full SVD. Again, see [4] for details.
Definition 10.28. Let A e C
nxn
and X e C
n
n
xn
. The transformation A i-> X
H
AX is called
a congruence. Note that a congruence is a similarity if and only ifX is unitary.
Note that congruence preserves the property of being Hermitian; i.e., if A is Hermitian,
then X
H
AX is also Hermitian. It is of interest to ask what other properties of a matrix are
preserved under congruence. It turns out that the principal property so preserved is the sign
of each eigenvalue.
Definition 10.29. Let A = A
H
e C"
x
" and let 7t, v, and denote the numbers of positive,
negative, and zero eigenvalues, respectively, of A. Then the inertia of A is the triple of
numbers In(A) = (n, v, ). The signature of A is given by sig(A) = n v.
Example 10.30.
2. If A = A" eC
n x
" , t h enA > 0 if and only if In (A) = (n, 0, 0).
3. If In(A) = (TT, v, ), then rank(A) = n + v.
Theorem 10.31 (Sylvester's Law of Inertia). Let A = A
H
e C
nxn
and X e C
n
n
xn
. Then
In(A) = ln(X
H
AX).
Proof: For the proof, see, for example, [21, p. 134]. D
Theorem 10.31 guarantees that rank and signature of a matrix are preserved under
congruence. We then have the following.
10.3. Equivalence Transformations and Congruence 103
Theorem 10.25 (Complete Orthogonal Decomposition). Let A E e ~ x n . Then there exist
unitary matrices U E e
mxm
and V E e
nxn
such that
where R E e;xr is upper (or lower) triangular with positive diagonal elements.
(10.5)
Theorem 10.26. Let A E e ~ x n . Then there exists a unitary matrix Q E e
mxm
and a
permutation matrix IT E en xn such that
QAIT = [ ~ ~ l
(10.6)
where R E e;xr is upper triangular and S E erx(n-r) is arbitrary but in general nonzero.
Remark 10.27. When A has full column rank but is "near" a rank deficient matrix,
various rank revealing QR decompositions are available that can sometimes detect such
phenomena at a cost considerably less than a full SVD. Again, see [4] for details.
Definition 10.28. Let A E e
nxn
and X E e ~ x n . The transformation A H- XH AX is called
a congruence. Note that a congruence is a similarity if and only if X is unitary.
Note that congruence preserves the property of being Hermitian; i.e., if A is Hermitian,
then X H AX is also Hermitian. It is of interest to ask what other properties of a matrix are
preserved under congruence. It turns out that the principal property so preserved is the sign
of each eigenvalue.
Definition 10.29. Let A = AH E e
nxn
and let rr, v, and ~ denote the numbers of positive,
negative, and zero eigenvalues, respectively, of A. Then the inertia of A is the triple of
numbers In(A) = (rr, v, n The signature of A is given by sig(A) = rr - v.
Example 10.30.
l.In[! 1
o 0
00] o 0
-10 =(2,1,1).
o 0
2. If A = AH E e
nxn
, then A> 0 if and only if In(A) = (n, 0, 0).
3. If In(A) = (rr, v, n, then rank(A) = rr + v.
Theorem 10.31 (Sylvester's Law of Inertia). Let A = A HE en xn and X E e ~ xn. Then
In(A) = In(X
H
AX).
Proof: For the proof, see, for example, [21, p. 134]. D
Theorem 10.31 guarantees that rank and signature of a matrix are preserved under
congruence. We then have the following.
H
e C"
xn
with In(A) = (jt, v, ). Then there exists a matrix
X e C"
n
xn
such that X
H
AX = diag(l, . . . , 1, -1,..., -1, 0, . . . , 0), where the number of
1 's is 7i, the number of l's is v, and the number 0/0 's is (,.
Proof: Let AI , . . . , X
w
denote the eigenvalues of A and order them such that the first TT are
positive, the next v are negative, and the final are 0. By Theorem 10.2 there exists a unitary
H
AU = diag(Ai, . . . , A
w
). Define the n x n matrix
Then it is easy to check that X = U W yields the desired result. D
Theorem 10.33. Suppose A = A
T
and D = D
T
. Then
if and only if either A > 0 and D - B
T
A~
l
B > 0, or D > 0 and A - BD^B
T
> 0.
Proof: The proof follows by considering, for example, the congruence
The details are straightforward and are left to the reader. D
Remark 10.34. Note the symmetric Schur complements of A (or D) in the theorem.
Theorem 10.35. Suppose A = A
T
and D = D
T
. Then
if and only ifA>0, AA
+
B = B, and D - B
T
A
+
B > 0.
Proof: Consider the congruence with
and proceed as in the proof of Theorem 10.33. D
10.4 Rational Canonical Form
One final canonical form to be mentioned is the rational canonical form.
Theorem 10.32. Let A = AH E c
nxn
with In(A) = (Jr, v, O. Then there exists a matrix
X E such that XH AX = diag(1, ... , I, -I, ... , -1,0, ... ,0), where the number of
1 's is Jr, the number of -I 's is v, and the numberofO's
Proof: Let A I, ... , An denote the eigenvalues of A and order them such that the first Jr are
positive, the next v are negative, and the final are O. By Theorem 10.2 there exists a unitary
matrix V such that VH AV = diag(AI, ... , An). Define the n x n matrix
vv = ... , 1/.f-Arr+I' ... , I/.f-Arr+v, I, ... ,1).
Then it is easy to check that X = V VV yields the desired result. 0
Theorem 10.33. Suppose A = AT and D = DT. Then
ifand only ifeither A> and D - BT A-I B > 0, or D > 0 and A - BD-
I
BT > O.
Proof: The proof follows by considering, for example, the congruence
B ] [I _A-I B JT [ A
D 0 I BT
] [
The details are straightforward and are left to the reader. 0
Remark 10.34. Note the symmetric Schur complements of A (or D) in the theorem.
Theorem 10.35. Suppose A = AT and D = DT. Then
B ] >
D -
if and only if A:::: 0, AA+B = B. and D - BT A+B:::: o.
Proof: Consider the congruence with
and proceed as in the proof of Theorem 10.33. 0
10.4 Rational Canonical Form
One final canonical form to be mentioned is the rational canonical form.
10.4. Rational Canonical Form 105
Definition 10.36. A matrix A e M"
x
" is said to be nonderogatory if its minimal polynomial
and characteristic polynomial are the same or, equivalently, if its Jordan canonical f orm
has only one block associated with each distinct eigenvalue.
Suppose A E W
xn
is a nonderogatory matrix and suppose its characteristic polyno-
mial is 7 r( A ) = A " ( a
0
+ A +
is similar to a matrix of the form
+ a
n
_ i A
n
~ ' ) - Then it can be shown (see [12]) that A
Definition 10.37. A matrix A e E
nx
" of the f orm (10.7) is called a companion matrix or
is said to be in companion form.
Companion matrices also appear in the literature in several equivalent forms. To
illustrate, consider the companion matrix
This matrix is a special case of a matrix in lower Hessenberg form. Using the reverse-order
identity similarity P given by (9.18), A is easily seen to be similar to the following matrix
in upper Hessenberg form:
Moreover, since a matrix is similar to its transpose (see exercise 13 in Chapter 9), the
following are also companion matrices similar to the above:
Notice that in all cases a companion matrix is nonsingular if and only if aO /= 0.
In fact, the inverse of a nonsingular companion matrix is again in companion form. For
*Yamr\1j=
10.4. Rational Canonical Form 105
Definition 10.36. A matrix A E lR
n
Xn is said to be nonderogatory ifits minimal polynomial
and characteristic polynomial are the same or; equivalently, if its Jordan canonical form
has only one block associated with each distinct eigenvalue.
Suppose A E lR
nxn
is a nonderogatory matrix and suppose its characteristic polyno-
mial is n(A) = An - (ao + alA + ... + an_IAn-I). Then it can be shown (see [12]) that A
is similar to a matrix of the form
o o o
o 0
o
(10.7)
o o
Definition 10.37. A matrix A E lR
nxn
of the form (10.7) is called a cornpanion rnatrix or
is said to be in cornpanion forrn.
Companion matrices also appear in the literature in several equivalent forms. To
illustrate, consider the companion matrix
(l0.8)
This matrix is a special case of a matrix in lower Hessenberg form. Using the reverse-order
identity similarity P given by (9.18), A is easily seen to be similar to the following matrix
in upper Hessenberg form:
a2 al
o 0
1 0
o 1
6]
o .
o
(10.9)
Moreover, since a matrix is similar to its transpose (see exercise 13 in Chapter 9), the
following are also companion matrices similar to the above:
l
:: ~ ! 0 1 ] .
ao 0 0
(10.10)
Notice that in all cases a companion matrix is nonsingular if and only if ao i= O.
In fact, the inverse of a nonsingular companion matrix is again in companion form. For
example,
o
1
o
- ~
ao
1
o
o
- ~
ao
o
o
_!!l
o
o
(10.11)
with a similar result for companion matrices of the form (10.10).
If a companion matrix of the form (10.7) is singular, i.e., if ao = 0, then its pseudo-
inverse can still be computed. Let a e M""
1
denote the vector \a\, 02,..., a
n
-i] and let
c =
l+
l
a
r
a
. Then it is easily verified that
Note that / caa
T
= (I + aa
T
) , and hence the pseudoinverse of a singular companion
matrix is not a companion matrix unless a = 0.
Companion matrices have many other interesting properties, among which, and per-
haps surprisingly, is the fact that their singular values can be found in closed form; see
[14].
Theorem 10.38. Let a\ > GI > > a
n
be the singular values of the companion matrix
A in (10.7). Let a = a\ + a\ + +a%_
{
and y = 1 + .Q + a. Then
Remark 10.39. Explicit formulas for all the associated right and left singular vectors can
also be derived easily.
If A R
nx
" is derogatory, i.e., has more than one Jordan block associated with
at least one eigenvalue, then it is not similar to a companion matrix of the form (10.7).
However, it can be shown that a derogatory matrix is similar to a block diagonal matrix,
each of whose diagonal blocks is a companion matrix. Such matrices are said to be in
rational canonical form (or Frobenius canonical form). For details, see, for example, [12].
Companion matrices appear frequently in the control and signal processing literature
but unfortunately they are often very difficult to work with numerically. Algorithms to reduce
an arbitrary matrix to companion form are numerically unstable. Moreover, companion
matrices are known to possess many undesirable numerical properties. For example, in
general and especially as n increases, their eigenstructure is extremely ill conditioned,
nonsingular ones are nearly singular, stable ones are nearly unstable, and so forth [14].
Ifao ^ 0, the largest and smallest singular values can also be written in the equivalent form
with a similar result for companion matrices of the form (10.10).
If a companion matrix of the form (10.7) is singular, i.e., if ao = 0, then its pseudo-
inverse can still be computed. Let a E JRn-1 denote the vector [ai, a2, ... , an-If and let
c = I + ~ T a' Then it is easily verified that
o
o
o
o
o o
o
o
+
o
1- caa
T
o J.
ca
Note that I - caa T = (I + aa T) -I , and hence the pseudoinverse of a singular companion
matrix is not a companion matrix unless a = O.
Companion matrices have many other interesting properties, among which, and per-
haps surprisingly, is the fact that their singular values can be found in closed form; see
[14].
Theorem 10.38. Let al ~ a2 ~ ... ~ an be the singular values of the companion matrix
A in (10.7). Leta = ar + ai + ... + a;_1 and y = 1 + aJ + a. Then
2 _ 1 ( J 2 2)
a
l
- 2 y + y - 4a
o
'
a? = 1 for i = 2, 3, ... , n - 1,
a; = ~ (y - J y2 - 4a
J
) .
If ao =1= 0, the largest and smallest singular values can also be written in the equivalent form
Remark 10.39. Explicit formulas for all the associated right and left singular vectors can
also be derived easily.
If A E JRnxn is derogatory, i.e., has more than one Jordan block associated with
at least one eigenvalue, then it is not similar to a companion matrix of the form (10.7).
However, it can be shown that a derogatory matrix is similar to a block diagonal matrix,
each of whose diagonal blocks is a companion matrix. Such matrices are said to be in
rational canonical form (or Frobenius canonical form). For details, see, for example, [12].
Companion matrices appear frequently in the control and signal processing literature
but unfortunately they are often very difficult to work with numerically. Algorithms to reduce
an arbitrary matrix to companion form are numerically unstable. Moreover, companion
matrices are known to possess many undesirable numerical properties. For example, in
general and especially as n increases, their eigenstructure is extremely ill conditioned,
nonsingular ones are nearly singular, stable ones are nearly unstable, and so forth [14].
Exercises 107
Companion matrices and rational canonical forms are generally to be avoided in floating-
point computation.
Remark 10.40. Theorem 10.38 yields some understanding of why difficult numerical
behavior might be expected for companion matrices. For example, when solving linear
systems of equations of the form (6.2), one measure of numerical sensitivity is K
P
(A) =
I I ^ I I
p
I I A~
l
I I
p
>
m
e so-called condition number of A with respect to inversion and with respect
to the matrix P-norm. I f this number is large, say 0(10*), one may lose up to k digits of
precision. I n the 2-norm, this condition number is the ratio of largest to smallest singular
values which, by the theorem, can be determined explicitly as
I t is easy to show that y/2/ao < k2(A) < --,, and when GO is small or y is large (or both),
then K2(A) ^ T~I. I t is not unusual for y to be large for large n. Note that explicit formulas
for K\ (A) and K oo(A) can also be determined easily by using (10.11).
EXERCISES
1. Show that if a triangular matrix is normal, then it must be diagonal.
2. Prove that if A e M"
x
" is normal, then Af(A) = A/"(A
r
).
3. Let A G C
nx
" and define p(A) = maxx
A(A) I ' M- Then p(A) is called the spectral

radius of A. Show that if A is normal, then p(A) = ||A||
2
. Show that the converse
is true if n = 2.
4. Let A C
nxn
be normal with eigenvalues y1 , ..., y
n
and singular values a\ > a
2
>
> o
n
> 0. Show that a, (A) = |A.,-(A)| for i e n.
5. Use the reverse-order identity matrix P introduced in (9.18) and the matrix U in
Theorem 10.5 to find a unitary matrix Q that reduces A e C"
x
" to lower triangular
form.
6. Let A = I J MeC
2x2
. Find a unitary matrix U such that
7. I f A e W
xn
is positive definite, show that A
[
must also be positive definite.
3. Suppose A e E"
x
" is positive definite. I s [ ^ /i 1 > 0?
}. Let R, S 6 E
nxn
be symmetric. Show that [ * J 1 > 0 if and only if S > 0 and
R> S
Exercises 107
Companion matrices and rational canonical forms are generally to be avoided in fioating-
point computation.
Remark 10.40. Theorem 10.38 yields some understanding of why difficult numerical
behavior might be expected for companion matrices. For example, when solving linear
systems of equations of the form (6.2), one measure of numerical sensitivity is Kp(A) =
II A II p II A -] II p' the so-called condition number of A with respect to inversion and with respect
to the matrix p-norm. If this number is large, say O(lO
k
), one may lose up to k digits of
precision. In the 2-norm, this condition number is the ratio of largest to smallest singular
values which, by the theorem, can be determined explicitly as
y+J
y
2- 4
a5
21
a
ol
It is easy to show that :::: K2(A) :::: 1:01' and when ao is small or y is large (or both),
then K2(A) It is not unusualfor y to be large forlarge n. Note that explicit formulas
for K] (A) and Koo(A) can also be determined easily by using (l0.11).
EXERCISES
1. Show that if a triangular matrix is normal, then it must be diagonal.
2. Prove that if A E jRnxn is normal, then N(A) = N(A
T
).
3. Let A E cc
nxn
and define peA) = max)..EA(A) IAI. Then peA) is called the spectral
radius of A. Show that if A is normal, then peA) = IIAII2' Show that the converse
is true if n = 2.
4. Let A E en xn be normal with eigenvalues A], ... , An and singular values 0'1 0'2
... an O. Show that a; (A) = IA;(A)I for i E!l.
5. Use the reverse-order identity matrix P introduced in (9.18) and the matrix U in
Theorem 10.5 to find a unitary matrix Q that reduces A E cc
nxn
to lower triangular
form.
6. Let A = :] E CC
2x2
. Find a unitary matrix U such that
7. If A E jRn xn is positive definite, show that A -I must also be positive definite.
8. Suppose A E jRnxn is positive definite. Is [1 O?
9. Let R, S E jRnxn be symmetric. Show that > 0 if and only if S > 0 and
R > S-I.
10. Find the inertia of the following matrices:
108
10. Find the inertia of the following matrices:
(a) [ ~ ~ l (b) [
(d) [-1 1 + j ]
1 - j -1 .
Chapter 10. Canonical Forms
-2 1 + j ]
1 - j -2 '
Chapter 11
Linear Differential and
Difference Equations
11.1 Differential Equations
In this section we study solutions of the linear homogeneous system of differential equations
for t > IQ. This is known as an initial-value problem. We restrict our attention in this
chapter only to the so-called time-invariant case, where the matrix A e R
nxn
is constant
and does not depend on t. The solution of (11.1) is then known always to exist and be
unique. It can be described conveniently in terms of the matrix exponential.
Definition 11.1. For all A e R
nxn
, the matrix exponential e
A
e R
nxn
is defined by the
power series
The series (11.2) can be shown to converge for all A (has radius of convergence equal
to +00). The solution of (11.1) involves the matrix
which thus also converges for all A and uniformly in t.
11.1.1 Properties of the matrix exponential
1. e = I.
Proof: This follows immediately from Definition 11.1 by setting A = 0.
2. For all A G R"
XM
, (e
A
f - e^.
Proof: This follows immediately from Definition 11.1 and linearity of the transpose.
109
Chapter 11
Linear Differential and
Difference Equations
11.1 Differential Equations
In this section we study solutions of the linear homogeneous system of differential equations
x(t) = Ax(t); x(to) = Xo E JR.n (11.1)
for t 2: to. This is known as an initial-value problem. We restrict our attention in this
chapter only to the so-called time-invariant case, where the matrix A E JR.nxn is constant
and does not depend on t. The solution of (11.1) is then known always to exist and be
unique. It can be described conveniently in terms of the matrix exponential.
Definition 11.1. For all A E JR.nxn, the matrix exponential e
A
E JR.nxn is defined by the
power series
+00 1
e
A
= L ,Ak.
k=O k.
(11.2)
The series (11.2) can be shown to converge for all A (has radius of convergence equal
to +(0). The solution of (11.1) involves the matrix
(11.3)
which thus also converges for all A and uniformly in t.
11.1.1 Properties of the matrix exponential
1. eO = I.
Proof This follows immediately from Definition 11.1 by setting A = O.
T T
2. For all A E JR.nxn, (e
A
) = e
A

Proof This follows immediately from Definition 11.1 and linearity of the transpose.
109
110 Chapter 11. Linear Differential and Difference Equations
3. For all A e R"
x
" and for all t, r e R, e
(t
+
T)A
= e'
A
e
rA
= e
lA
e'
A
.
Proof: Note that
Compare like powers of A in the above two equations and use the binomial theorem
on (t + T)*.
4. For all A, B e R"
xn
and for all t e R, e
t(A+B)
=ê'
A
e'
B
= e'
B
e'
A
if and only if A
and B commute, i.e., AB = B A.
Proof: Note that
and
and
while
Compare like powers of t in the first equation and the second or third and use the
binomial theorem on (A + B)
k
and the commutativity of A and B.
5. For all A e R"
x
" and for all t e R, (e'
A
)~
l
= e~'
A
.
Proof: Simply take T = t in property 3.
6. Let denote the Laplace transform and ~
!
the inverse Laplace transform. Then for
all A R"
x
" and for all t R,
(a) C{e
tA
} = (sI-Ar
l
.
(b) -
1
{(j/-A)-
1
} =
M
.
Proof: We prove only (a). Part (b) follows similarly.
3. For all A E JRnxn and for all t, T E JR, e(t+r)A = etA erA = erAe
tA
.
Proof" Note that
(t + T)2 2
e(t+r)A = I + (t + T)A + A + ...
2!
and
tA rA t 2 T 2
(
2 )( 2 )
e e = I + t A + 2! A +... I + T A + 2! A +... .
Compare like powers of A in the above two equations and use the binomial theorem
on(t+T)k.
4. For all A, B E JRnxn and for all t E JR, et(A+B) =-etAe
tB
= etBe
tA
if and only if A
and B commute, i.e., AB = BA.
Proof' Note that
and
while
t
2
et(A+B) = I + teA + B) + -(A + B)2 + ...
2!
tB tA t 2 t 2
(
2 )( 2 )
e e = 1+ tB + 2iB +... 1+ tA + 2!A +... .
Compare like powers of t in the first equation and the second or third and use the
binomial theorem on (A + B/ and the commutativity of A and B.
5. ForaH A E JRnxn and for all t E JR, (etA)-1 = e-
tA
.
Proof" Simply take T = -t in property 3.
6. Let denote the Laplace transform and -1 the inverse Laplace transform. Then for
all A E JRnxn and for all t E lR,
(a) .l{e
tA
} = (sI - A)-I.
(b) .l-I{(sl- A)-I} = erA.
Proof" We prove only (a). Part (b) follows similarly.
{+oo
= io et(-sl)e
tA
dt
(+oo
= io ef(A-sl) dt since A and (-sf) commute
11.1. Differential Equations 111
= (sl -A)-
1
.
The matrix (s I A) ~' is called the resolvent of A and is defined for all s not in A (A).
Notice in the proof that we have assumed, for convenience, that A is diagonalizable.
If this is not the case, the scalar dyadic decomposition can be replaced by
using the JCF. All succeeding steps in the proof then follow in a straightforward way.
7. For all A e R"
x
" and for all t e R, (e'
A
) = Ae
tA
= e'
A
A.
Proof: Since the series (11.3) is uniformly convergent, it can be differentiated term-by-
term from which the result follows immediately. Alternatively, the formal definition
can be employed as follows. For any consistent matrix norm,
= {+oo t e(Ai-S)t x;y;H dt assuming A is diagonalizable
10 ;=1
= e(Ai-S)t dt]x;y;H
n 1
= '"' --Xi y;H assuming Re s > Re Ai for i E !!
L..... s - A"
i=1 I
= (sI - A)-I.
The matrix (s I - A) -I is called the resolvent of A and is defined for all s not in A (A).
Notice in the proof that we have assumed, for convenience, that A is diagonalizable.
If this is not the case, the scalar dyadic decomposition can be replaced by
m
et(A-sl) = L Xiet(Ji-sl)y;H
;=1
using the JCF. All succeeding steps in the proof then follow in a straightforward way.
7. For all A E JRnxn and for all t E JR, 1h(e
tA
) = Ae
tA
= etA A.
Proof: Since the series (11.3) is uniformly convergent, it can be differentiated term-by-
term from which the result follows immediately. Alternatively, the formal definition
d e(t+M)A _ etA
_(/A) = lim
d t L'lt
can be employed as follows. For any consistent matrix norm,
II
etA II III II
---u.--- - Ae
tA
= L'lt - /A) - Ae
tA
= II - etA) - Ae
tA
II
= II - l)e
tA
- Ae
tA
II
II
I ( (M)2 2 ) tA tAil
= L'lt M A + A +... e - Ae
= II ( Ae
tA
+ A
2
e
tA
+ ... ) - Ae
tA
II
= II ( A2 + A
3
+ .. , ) etA II
< MIIA21111e
tA
II _ + -IIAII + --IIAI12 + ...
(
1 L'lt (L'lt)2 )
- 2! 3! 4!
< L'lt1lA21111e
tA
Il (1 + L'ltiIAIl + IIAII2 + ... )
= L'lt IIA 21111e
tA

For fixed t, the right-hand side above clearly goes to 0 as At goes to 0. Thus, the
limit exists and equals Ae'
A
. A similar proof yields the limit e'
A
A, or one can use the
fact that A commutes with any polynomial of A of finite degree and hence with e'
A
.
nxn
. The solution of the linear homogeneous initial-value problem
Proof: Differentiate (11.5) and use property 7 of the matrix exponential to get x( t ) =
Ae
( t
~
to) A
xo = Ax( t) . Also, x( t
0
) e
( fo
~
t
'
) A
X Q X Q so, by the fundamental existence and
uniqueness theorem for ordinary differential equations, (11.5) is the solution of (11.4). D
nxn
, B e W
xm
and let the vector-valued function u be given
and, say, continuous. Then the solution of the linear inhomogeneous initial-value problem
for t > IQ is given by the variation of parameters formula
Proof: Differentiate (11.7) and again use property 7 of the matrix exponential. The general
formula
is used to get x( t ) = Ae
{
'-
to) A
x
0
+ f'
o
Ae
(
'-
s) A
Bu( s) ds + Bu( t) = Ax( t) + Bu( t) . Also,
*('o)
=
< ?
(f
~
fo)/ 1
. o + 0 = X Q so, by the fundamental existence and uniqueness theorem for
ordinary differential equations, (11.7) is the solution of (11.6). D
Remark 11.4. The proof above simply verifies the variation of parameters formula by
direct differentiation. The formula can be derived by means of an integrating factor "trick"
as follows. Premultiply the equation x Ax = Bu by e~
tA
to get
For fixed t, the right-hand side above clearly goes to 0 as t:.t goes to O. Thus, the
limit exists and equals Ae
t
A A similar proof yields the limit e
t
A A, or one can use the
fact that A commutes with any polynomial of A of finite degree and hence with etA.
Theorem 11.2. Let A E IR
n
xn. The solution of the linear homogeneous initial-value problem
x(t) = Ax(l); x(to) = Xo E IR
n
(11.4)
for t ::: to is given by
(11.5)
Proof: Differentiate (11.5) and use property 7 of the matrix exponential to get x (t) =
Ae(t-to)A
xo
= Ax(t). Also, x(to) = e(to-to)A Xo = Xo so, by the fundamental existence and
uniqueness theorem for ordinary differential equations, (11.5) is the solution of (11.4). 0
nxn
, B E IR
nxm
and let the vector-valued function u be given
and, say, continuous. Then the solution of the linear inhomogeneous initial-value problem
x(t) = Ax(t) + Bu(t); x(to) = Xo E IR
n
for t ::: to is given by the variation of parameters formula
x(t) = e(t-to)A
xo
+ t e(t-s)A Bu(s) ds.
l t o
(11.6)
(11.7)
Proof: Differentiate (11.7) and again use property 7 of the matrix exponential. The general
formula
d l
q
(t) l
q
(t) af(x t) dq(t) dp(t)
- f(x, t) dx = ' dx + f(q(t), t)-- - f(p(t), t)--
dt pet) pet) at dt dt
is used to get x(t) = Ae(t-to)A Xo + Ir: Ae(t-s)A Bu(s) ds + Bu(t) = Ax(t) + Bu(t). Also,
x(t
o
} = e(to-tolA Xo + 0 = Xo so, by the fundilm()ntill nnd uniqu()Oc:s:s theorem for
ordinary differential equations, (11.7) is the solution of (1l.6). 0
Remark 11.4. The proof above simply verifies the variation of parameters formula by
direct differentiation. The formula can be derived by means of an integrating factor "trick"
as follows. Premultiply the equation x - Ax = Bu by e-
tA
to get
(11.8)
Now integrate (11.8) over the interval [to, t]:
Thus,
and hence
11.1.4 Linear matrix differential equations
Matrix-valued initial-value problems also occur frequently. The first is an obvious general-
ization of Theorem 11.2, and the proof is essentially the same.
Theorem 11.5. Let A e W
lxn
. The solution of the matrix linear homogeneous initial-value
nrohlcm
for t > to is given by
In the matrix case, we can have coefficient matrices on both the right and left. For
convenience, the following theorem is stated with initial time to = 0.
Theorem 11.6. Let A e Rn
xn
, B e R
mxm
, and C e Rn
xm
. Then the matrix initial-value
problem
a
tA
ra
tB
has the solutionX ( t ) = e Ce
Proof: Differentiate e
tA
Ce
tB
with respect to t and use property 7 of the matrix exponential.
The fact that X ( t ) satisfies the initial condition is trivial. D
Corollary 11.7. Let A, C e IR"
X
". Then the matrix initial-value problem
has the solution X(t} = e
tA
Ce
tAT
.
When C is symmetric in (11.12), X ( t ) is symmetric and (11.12) is known as a Lya-
punov differential equation. The initial-value problem (11.11) is known as a Sylvester
differential equation.
11.1. Differential Equations
Now integrate (11.8) over the interval [to, t]:
Thus,
and hence
-e-sAx(s) ds = e-SABu(s) ds.
1
t d 1t
to ds to
e-tAx(t) - e-toAx(to) = t e-
sA
Bu(s) ds
lto
x(t) = e(t-t
olA
xo
+ t e(t-s)A Bu(s) ds.
lto
11.1.4 Linear matrix differential equations
113
Matrix-valued initial-value problems also occur frequently. The first is an obvious general-
ization of Theorem 11.2, and the proof is essentially the same.
Theorem 11.5. Let A E jRnxn. The solution of the matrix linear homogeneous initial-value
problem
X(t) = AX(t); X(to) = C E jRnxn (11.9)
for t ::: to is given by
X(t) = e(t-to)Ac.
(11.10)
In the matrix case, we can have coefficient matrices on both the right and left. For
convenience, the following theorem is stated with initial time to = O.
Theorem 11.6. Let A E jRnxn, B E jRmxm, and C E ]R.nxm. Then the matrix initial-value
problem
X(t) = AX(t) + X(t)B; X(O) = C (11.11)
has the solution X(t) = etACe
tB
.
Proof: Differentiate etACe
tB
with respect to t and use property 7 of the matrix exponential.
The fact that X (t) satisfies the initial condition is trivial. 0
Corollary 11.7. Let A, C E ]R.nxn. Then the matrix initial-value problem
X(t) = AX(t) + X(t)AT; X(O) = C (11.12)
has the solution X(t) = etACetAT.
When C is symmetric in (11.12), X (t) is symmetric and (11.12) is known as a Lya-
punov differential equation. The initial-value problem (11.11) is known as a Sylvester
differential equation.
11.1.5 Modal decompositions
Let A E W
xn
and suppose, for convenience, that it is diagonalizable (if A is not diagonaliz-
able, the rest of this subsection is easily generalized by using the JCF and the decomposition
A ^ X f Ji Y
t
H
as discussed in Chapter 9). Then the solution x(t) of (11.4) can be written
The ki s are called the modal velocities and the right eigenvectors *, are called the modal
directions. The decomposition above expresses the solution x(t) as a weighted sum of its
modal velocities and directions.
This modal decomposition can be expressed in a different looking but identical form
if we write the initial condition X Q as a weighted sum of the right eigenvectors
Then
In the last equality we have used the fact that y f * X j = S f j .
Similarly, in the inhomogeneous case we can write
JCF method
Let A e R"
x
" and suppose X e Rn
xn
is such that X"
1
AX = J, where J is a JCF for A.
Then
11.1 .5 Modal decompositions
Let A E jRnxn and suppose, for convenience, that it is diagonalizable (if A is not diagonaliz-
able, the rest of this subsection is easily generalized by using the JCF and the decomposition
A = L X;li y
i
H
as discussed in Chapter 9). Then the solution x(t) of (11.4) can be written
x(t) = e(t-to)A Xo
= (ti.iU-tO)Xiyr) Xo
1=1
n
= L(Yi
H
xoeAi(t-tOXi.
i=1
The Ai s are called the modal velocities and the right eigenvectors Xi are called the modal
directions. The decomposition above expresses the solution x (t) as a weighted sum of its
modal velocities and directions.
This modal decomposition can be expressed in a different looking but identical form
n
if we write the initial condition Xo as a weighted sum of the right eigenvectors Xo = L ai Xi.
Then
n
= L(aieAiU-tOXi.
i=1
In the last equality we have used the fact that Yi
H
X j = flij.
Similarly, in the inhomogeneous case we can write
i
t e(t-s)A Bu(s) ds = t (it eAiU-S)YiH Bu(s) dS) Xi.
~ i=1 ~
JCF method
i=1
Let A E jRnxn and suppose X E j R ~ x n is such that X-I AX = J, where J is a JCF for A.
Then
etA = etXJX-1
= XetJX-
1
I
n
Le
A
,
X'Yi
H
if A is diagonalizable
1=1
~ t,x;e'J,y;H in geneml.
If A is diagonalizable, it is then easy to compute e
tA
via the formula e
tA
= Xe
tJ
X '
since e
tj
is simply a diagonal matrix.
In the more general case, the problem clearly reduces simply to the computation of
the exponential of a Jordan block. To be specific, let .7, e <C
kxk
be a Jordan block of the form
Clearly A/ and N commute. Thus, e
tJi
= e'
u
e
tN
by property 4 of the matrix exponential.
The diagonal part is easy: e
tu
= diag(e
x
',..., e
xt
}. But e
tN
is almost as easy since N is
nilpotent of degree k.
Definition 11.8. A matrix M e M
nx
"
M
p
= 0, while M
p
~
l
^ 0.
is nilpotent of degree (or index, or grade) p if
For the matrix N defined above, it is easy to check that while N has 1's along only
its first superdiagonal (and O's elsewhere), N
2
has 1's along only its second superdiagonal,
and so forth. Finally, N
k
~
l
has a 1 in its (1, k) element and has O's everywhere else, and
N
k
= 0. Thus, the series expansion of e'
N
is finite, i.e.,
Thus,
In the case when A. is complex, a real version of the above can be worked out.
If A is diagonalizable, it is then easy to compute etA via the formula etA = Xe
tl
X-I
since e
t
I is simply a diagonal matrix.
In the more general case, the problem clearly reduces simply to the computation of
the exponential of a Jordan block. To be specific, let J
i
E C
kxk
J
i
=
A 1
o A
o
o o
o =U+N.
o A
Clearly AI and N commute. Thus, e
t
I, = eO.! e
l
N by property 4 of the matrix exponential.
The diagonal part is easy: e
lH
= diag(e
At
, ,eAt). But e
lN
is almost as easy since N is
nilpotent of degree k.
Definition 11.8. A matrix M E jRnxn is nilpotent of degree (or index, or grade) p if
MP = 0, while MP-I t=- O.
For the matrix N defined above, it is easy to check that while N has l's along only
its first superdiagonal (and O's elsewhere), N
2
has l's along only its second superdiagonal,
and so forth. Finally, N
k
-
I
has a 1 in its (1, k) element and has O's everywhere else, and
N
k
= O. Thus, the series expansion of e
lN
is finite, i.e.,
Thus,
t
2
t
k
-
I
e
IN
=I+tN+-N
2
+ ... + N
k
-
I
2! (k - I)!
o
o o
eAt
teAt
12 At
2I
e
0
eAt teAl
ell; =
0 0
eAt
0 0
t
1
Ik-I At
(k-I)! e
12 At
2I
e
teAl
eAt
In the case when A is complex, a real version of the above can be worked out.
Example 11.9. Let A = [ ~ _ \ J]. Then A (A) = {-2, -2} and
Interpolation method
This method is numerically unstable in finite-precision arithmetic but is quite effective for
hand calculation in small-order problems. The method is stated and illustrated for the
exponential function but applies equally well to other functions.
Given A E.
nxn
and /(A) = e
tx
, compute f(A) = e'
A
, where t is a fixed scalar.
Suppose the characteristic polynomial of A can be written as n ( X ) = Yi?=i (^ ~~ ^ i)" '
where the A., - s are distinct. Define
where O TQ , . . . , a
n
-i are n constants that are to be determined. They are, in fact, the unique
solution of the n equations:
Here, the superscript (&) denotes the fcth derivative with respect to X. With the a, s then
known, the function g is known and /(A) = g(A). The motivation for this method is
the Cayley-Hamilton Theorem, Theorem 9.3, which says that all powers of A greater than
n 1 can be expressed as linear combinations of A
k
for k = 0, 1, . . . , n 1. Thus, all the
terms of order greater than n 1 in the power series for e'
A
can be written in terms of these
lower-order powers as well. The polynomial g gives the appropriate linear combination.
Example 11.10. Let
and /(A) = e
tK
. Then j r(A.) = -(A. + I)
3
, so m = 1 and n
{
= 3.
Let g(X) UQ + a\X + oÂ.
2
. Then the three equations for the a, s are given by
Example 11.9.
Let A = [=i
a Then A(A) = {-2, -2} and
etA = Xe
tJ
x-I
=[
2 1
] exp t [
-2
- ~ ] [
-1
]
0
-1 2
=[
2
] [ e ~ 2 t
te-
2t
] [
-1
]
1
e-
2t
-1 2
Interpolation method
This method is numerically unstable in finite-precision arithmetic but is quite effective for
hand calculation in small-order problems. The method is stated and illustrated for the
exponential function but applies equally well to other functions.
Given A E jRnxn and f(A) = etA, compute f(A) = etA, where t is a fixed scalar.
Suppose the characteristic polynomial of A can be written as n(A) = nr=1 (A - Ai t',
where the Ai s are distinct. Define
where ao, ... , an-l are n constants that are to be determined. They are, in fact, the unique
solution of the n equations:
g(k)(Ai) = f(k)(Ai); k = 0, I, ... , ni - I, i Em.
Here, the superscript (k) denotes the kth derivative with respect to A. With the aiS then
known, the function g is known and f(A) = g(A). The motivation for this method is
the Cayley-Hamilton Theorem, Theorem 9.3, which says that all powers of A greater than
n - 1 can be expressed as linear combinations of A k for k = 0, I, ... , n - 1. Thus, all the
terms of order greater than n - 1 in the power series for e
t
A can be written in terms of these
lower-order powers as well. The polynomial g gives the appropriate linear combination.
Example 11.10. Let
A = - ~ - ~ ~ ]
o 0-1
and f(A) = etA. Then n(A) = -(A + 1)3, so m = 1 and nl = 3.
Let g(A) = ao + alA + a2A2. Then the three equations for the aiS are given by
g(-I) = f(-1) ==> ao -al +a2 = e-
t
,
g'(-1) = f'(-1) ==> at - 2a2 = te-
t
,
g"(-I) = 1"(-1) ==> 2a2 = t
2
e-
t

Solving for the a, s, we find
Thus,
~4 4i t f f > \ t k TU^^ _/"i\ f \ i o\ 2
Example 11.11. Let A = [ _* J] and /(A) = e
a
. Then 7 r(X ) = (A + 2)
2
so m = 1 and
i = 2.
Let g(A.) = o + ofiA.. Then the defining equations for the a,-s are given by
Solving for the a,s, we find
Thus,
Other methods
1. Use e
tA
= ~
l
{(sl A)^
1
} and techniques for inverse Laplace transforms. This
is quite effective for small-order problems, but general nonsymbolic computational
techniques are numerically unstable since the problem is theoretically equivalent to
knowing precisely a JCF.
2. Use Pade approximation. There is an extensive literature on approximating cer-
tain nonlinear functions by rational functions. The matrix analogue yields e
A
=
11 .1. Differential Equations
117
Solving for the ai s, we find
Thus,
Example 11.11. Let A = : : : : ~ 6] and f(A) = eO-. Then rr(A) = (A + 2)2 so m = 1 and
nL = 2.
Let g(A) = ao + aLA. Then the defining equations for the aiS are given by
g(-2) = f(-2) ==> ao - 2al = e-
2t
,
g'(-2) = f'(-2) ==> al = te-
2t
.
Solving for the aiS, we find
Thus,
ao = e-
2t
+ 2te-
2t
,
aL = te-
2t
.
f(A) = etA = g(A) = aoI + al A
= (e-
2t
+ 2te-
2t
) [ ~
_ [ e-
2t
_ 2te-
2t
- -te-
2t
Other methods
o ] + te-
2t
[-4 4 ]
I -I 0
1. Use etA = .c-I{(sI - A)-I} and techniques for inverse Laplace transforms. This
is quite effective for small-order problems, but general nonsymbolic computational
techniques are numerically unstable since the problem is theoretically equivalent to
knowing precisely a JCE
2. Use Pade approximation. There is an extensive literature on approximating cer-
tain nonlinear functions by rational functions. The matrix analogue yields e
A
~
D~
l
(A)N(A), where D(A) = 8
0
I + Si A H h S
P
A
P
and N(A) = v
0
I + v
l
A +
+ v
q
A
q
. Explicit formulas are known for the coefficients of the numerator and
denominator polynomials of various orders. Unfortunately, a Fade approximation for
the exponential is accurate only in a neighborhood of the origin; in the matrix case
this means when || A|| is sufficiently small. This can be arranged by scaling A, say, by
/ * \
2
*
multiplying it by 1/2* for sufficiently large k and using the fact that e
A
= ( e
{ ] / 2 )A
j .
Numerical loss of accuracy can occur in this procedure from the successive squarings.
3. Reduce A to (real) Schur form S via the unitary similarity U and use e
A
= Ue
s
U
H
and successive recursions up the superdiagonals of the (quasi) upper triangular matrix
e
s
.
4. Many methods are outlined in, for example, [19]. Reliable and efficient computation
of matrix functions such as e
A
and log(A) remains a fertile area for research.
11.2 Difference Equations
In this section we outline solutions of discrete-time analogues of the linear differential
equations of the previous section. Linear discrete-time systems, modeled by systems of
difference equations, exhibit many parallels to the continuous-time differential equation
case, and this observation is exploited frequently.
xn
. The solution of the linear homogeneous system of difference
equations
nxn
, B e R
nxm
and suppose {*} a given sequence of
m-vectors. Then the solution of the inhomogeneous initial-value problem
for k > 0 is given by
Proof: The proof is almost immediate upon substitution of (11.14) into (11.13). D
Remark 11.13. Again, we restrict our attention only to the so-called time-invariant
case, where the matrix A in (11.13) is constant and does not depend on k. We could also
consider an arbitrary "initial time" ko, but since the system is time-invariant, and since we
want to keep the formulas "clean" (i.e., no double subscripts), we have chosen ko = 0 for
convenience.
D-I(A)N(A), where D(A) = 001 + olA + ... + opAP and N(A) = vol + vIA +
... + Vq A q. Explicit formulas are known for the coefficients of the numerator and
denominator polynomials of various orders. Unfortunately, a Pad6 approximation for
the exponential is accurate only in a neighborhood of the origin; in the matrix case
this means when IIAII is sufficiently small. This can be arranged by scaling A, say, by
2'
multiplying it by 1/2k for sufficiently large k and using the fact that e
A
= (e( I /2')A )
Numerical loss of accuracy can occur in this procedure from the successive squarings.
3. Reduce A to (real) Schur form S via the unitary similarity U and use e
A
= U e
S
U H
and successive recursions up the superdiagonals of the (quasi) upper triangular matrix
e
S
.
4. Many methods are outlined in, for example, [19]. Reliable and efficient computation
of matrix functions such as e
A
and 10g(A) remains a fertile area for research.
11.2 Difference Equations
In this section we outline solutions of discrete-time analogues of the linear differential
equations of the previous section. Linear discrete-time systems, modeled by systems of
difference equations, exhibit many parallels to the continuous-time differential equation
case, and this observation is exploited frequently.
Theorem 11.12. Let A E jRn xn. The solution of the linear homogeneous system of difference
equations
(11.13)
for k 2:: 0 is given by
Proof: The proof is almost immediate upon substitution of (11.14) into (11.13). 0
Remark 11.13. Again, we restrict our attention only to the so-called time-invariant
case, where the matrix A in (11.13) is constant and does not depend on k. We could also
consider an arbitrary "initial time" ko, but since the system is time-invariant, and since we
want to keep the formulas "clean" (i.e., no double subscripts), we have chosen ko = 0 for
convenience.
Theorem 11.14. Let A E jRnxn, B E jRnxm and suppose { u d t ~ is a given sequence of
m-vectors. Then the solution of the inhomogeneous initial-value problem
(11.15)
11.2. Difference Equations 119
is given by
11.2.3 Computation of matrix powers
It is clear that solution of linear systems of difference equations involves computation of
A
k
. One solution method, which is numerically unstable but sometimes useful for hand
calculation, is to use z-transforms, by analogy with the use of Laplace transforms to compute
a matrix exponential. One definition of the z-transform of a sequence {gk} is
Assuming |z| > max |A|, the z-transform of the sequence {A
k
} is then given by
XA(A)
Proof: The proof is again almost immediate upon substitution of (11.16) into (11.15). D
Methods based on the JCF are sometimes useful, again mostly for small-order prob-
lems. Assume that A e M"
xn
and let X e R^
n
be such that X~
1
AX = /, where J is a
JCF for A. Then
If A is diagonalizable, it is then easy to compute A
k
via the formula A
k
XJ
k
X
l
since /* is simply a diagonal matrix.
11.2. Difference Equations
is given by
k-I
xk=AkXO+LAk-j-IBUj, k:::.O.
j=O
119
(11.16)
Proof: The proof is again almost immediate upon substitution of (11.16) into (11.15). 0
11.2.3 Computation of matrix powers
It is clear that solution of linear systems of difference equations involves computation of
A k. One solution method, which is numerically unstable but sometimes useful for hand
calculation, is to use z-transforms, by analogy with the use of Laplace transforms to compute
a matrix exponential. One definition of the z-transform of a sequence {gk} is
+00
= LgkZ-
k
.
k=O
Assuming Izl > max IAI, the z-transform of the sequence {Ak} is then given by
AEA(A)
+00
k "'kk 1 12
Z({A})=L...-z-A =I+-A+"2A + ...
k=O z z
= (l-z-IA)-I
= z(zI - A)-I.
Methods based on the JCF are sometimes useful, again mostly for small-order prob-
lems. Assume that A E jRnxn and let X E be such that X-I AX = J, where J is a
JCF for A. Then
Ak = (XJX-I)k
= XJkX-
1
_I
- m
LXi Jty
i
H
;=1
if A is diagonalizable,
in general.
If A is diagonalizable, it is then easy to compute Ak via the formula Ak = X Jk X-I
since Jk is simply a diagonal matrix.
In the general case, the problem again reduces to the computation of the power of a
Jordan block. To be specific, let 7, e C
pxp
Writing /, = XI + N and noting that XI and the nilpotent matrix N commute, it is
then straightforward to apply the binomial theorem to (XI + N)
k
and verify that
The symbol ( ) has the usual definition of ,
(
^ ., and is to be interpreted as 0 if k < q.
In the case when A. is complex, a real version of the above can be worked out.
-4
Example 11.15. Let A = [_J J]. Then
Basic analogues of other methods such as those mentioned in Section 11.1.6 can also
be derived for the computation of matrix powers, but again no universally "best" method
exists. For an erudite discussion of the state of the art, see [11, Ch. 18].
11.3 Higher-Order Equations
It is well known that a higher-order (scalar) linear differential equation can be converted to
a first-order linear system. Consider, for example, the initial-value problem
with 4 > (t } a given function and n initial conditions
In the general case, the problem again reduces to the computation of the power of a
Jordan block. To be specific, let J
i
E Cpxp be a Jordan block of the form
o ... 0 A
Writing J
i
= AI + N and noting that AI and the nilpotent matrix N commute, it is
then straightforward to apply the binomial theorem to (AI + N)k and verify that
Ak
kA k-I
(;)A
k
-
2
(
k ) Ak-P+I
p-l
0
Ak kA
k
-
1
J/ =
0 0
Ak
( ; ) A
k
-
2
kA
k
-
1
0 0
Ak
The symbol (: ) has the usual definition of ! ( k k ~ q ) ! and is to be interpreted as 0 if k < q.
In the case when A is complex, a real version of the above can be worked out.
Example 11.15. Let A = [=i a Then
Ak = XJkX-1 = [2 1 ] [(_2)k k(-2)kk-
1
] [ 1 -2
1
]
1 1 0 (-2) -1
_ [ (_2/-
1
(-2 - 2k) k( -2l+
1
]
- -k( _2)k-1 (-2l-
1
(2k - 2) .
Basic analogues of other methods such as those mentioned in Section 11.1.6 can also
be derived for the computation of matrix powers, but again no universally "best" method
exists. For an erudite discussion of the state of the art, see [11, Ch. 18].
11.3 Higher-Order Equations
It is well known that a higher-order (scalar) linear differential equation can be converted to
a first-order linear system. Consider, for example, the initial-value problem
(11.17)
with J(t) a given function and n initial conditions
y(O) = Co, y(O) = CI, ... , in-I)(O) = Cn-I' (1l.l8)
Exercises 121
Here, v
(m)
denotes the mth derivative of y with respect to t. Define a vector x (?) e R" with
components *i(0 = y ( t ) , x
2
( t) = y ( t ) , . . . , x
n
( t) = y
{ n
~
l )
( t ) . Then
These equations can then be rewritten as the first-order linear system
The initial conditions take the form ^(0) = c = [ C Q , c\, ..., C
M
_ I ] .
Note that det(X7 A) = A." + a
n
-\X
n
~
l
H h a\X + ao. However, the companion
matrix A in (11.19) possesses many nasty numerical properties for even moderately sized n
and, as mentioned before, is often well worth avoiding, at least for computational purposes.
A similar procedure holds for the conversion of a higher-order difference equation
EXERCISES
1. Let P R
nxn
be a projection. Show that e
p
% / + 1.718P.
2. Suppose x, y R" and let A = xy
T
. Further, let a = x
T
y. Show that e'
A
I + g ( t , a) xy
T
, where
3. Let
with n initial conditions, into a linear first-order difference equation with (vector) initial
condition.
Exercises 121
Here, y(m) denotes the mth derivative of y with respect to t. Define a vector x (t) E ]Rn with
components Xl (t) = yet), X2(t) = yet), ... , Xn(t) = In-l)(t). Then
Xl (I) = X2(t) = y(t),
X2(t) = X3(t) = yet),
Xn-l (t) = Xn(t) = y(n-l)(t),
Xn(t) = y(n)(t) = -aoy(t) - aly(t) - ... - an_lln-l)(t) + (t)
= -aOx\ (t) - a\X2(t) - ... - an-lXn(t) + (t).
These equations can then be rewritten as the first-order linear system
0 0 0
0 0 1
x(t)+ [ n ( t )
x(t) =
0
0 0 1
-ao -a\ -a
n
-\
The initial conditions take the form X (0) = C = [co, Cl, .. , C
n
-\ r.
(11.19)
Note that det(A! - A) = An + an_1A
n
-
1
+ ... + alA + ao. However, the companion
matrix A in (11.19) possesses many nasty numerical properties for even moderately sized n
and, as mentioned before, is often well worth avoiding, at least for computational purposes.
A similar procedure holds for the conversion of a higher-order difference equation
with n initial conditions, into a linear first-order difference equation with (vector) initial
condition.
EXERCISES
1. Let P E lR
nxn
be a projection. Show that e
P
~ ! + 1.718P.
2. Suppose x, y E lR
n
and let A = xyT. Further, let a = XT y. Show that etA
1+ get, a)xyT, where
{
!(eat - I)
g(t,a)= a t
3. Let
if a 1= 0,
if a = O.
122 Chapter 11. L i n ear Di f f eren ti al and Di f f erence Equati on s
where X e M'
nx
" is arbitrary. Show that
4. Let K denote the skew-symmetric matrix
where / denotes the n x n identity matrix. A matrix A e R
2n x2n
is said to be
Hamiltonian if K~
1
A
T
K = -A and to be symplectic if K~
l
A
T
K - A-
1
.
(a) Suppose E is Hamiltonian and let A, be an eigenvalue of H. Show that A, must
also be an eigenvalue of H.
(b) Suppose S is symplectic and let A. be an eigenvalue of S. Show that 1 /A, must
also be an eigenvalue of S.
(c) Suppose that H is Hamiltonian and S is symplectic. Show that S~
1
HS must be
Hamiltonian.
(d) Suppose H is Hamiltonian. Show that e
H
must be symplectic.
5. Let a, ft R and
Then show that
6. Find a general expression for
7. Find e
M
when A =
5. Let
(a) Solve the differential equation
where X E jRmxn is arbitrary. Show that
e
A = [eo I sinh 1 X ]
~ I .
4. Let K denote the skew-symmetric matrix
[
0 In ]
-In 0 '
where In denotes the n x n identity matrix. A matrix A E jR2nx2n is said to be
Hamiltonian if K -I AT K = - A and to be symplectic if K -I AT K = A -I.
(a) Suppose H is Hamiltonian and let).. be an eigenvalue of H. Show that -).. must
also be an eigenvalue of H.
(b) Suppose S is symplectic and let).. be an eigenvalue of S. Show that 1/).. must
also be an eigenValue of S.
(c) Suppose that H is Hamiltonian and S is symplectic. Show that S-I H S must be
Hamiltonian.
(d) Suppose H is Hamiltonian. Show that e
H
must be symplectic.
5. Let a, f3 E lR and
Then show that
6. Find a general expression for
7. Find etA when A =
8. Let
ectt cos f3t
_eut sin f3t
ectctrt sin t J.
e cos/A
(a) Solve the differential equation
i = Ax ; x(O) = [ ~ J.
Exercises 123
Show that the eigenvalues of the solution X ( t ) of this problem are the same as those
of Cf or all?.
11. The year is 2004 and there are three large "free trade zones" in the world: Asia (A),
Europe (E), and the Americas (R). Suppose certain multinational companies have
total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R. Each
year half of the Americas' money stays home, a quarter goes to Europe, and a quarter
goes to Asia. For Europe and Asia, half stays home and half goes to the Americas.
(a) Find the matrix M that gives
(b) Find the eigenvalues and right eigenvectors of M.
(c) Find the distribution of the companies' assets at year k.
(d) Find the limiting distribution of the $40 trillion as the universe ends, i.e., as
k +00 (i.e., around the time the Cubs win a World Series).
(Exercise adapted from Problem 5.3.11 in [24].)
(b) Solve the differential equation
9. Consider the initial-value problem
for t > 0. Suppose that A e E"
x
" is skew-symmetric and let a = \\XQ\\
2
. Show that
||*(OII
2
= af or al l f > 0.
10. Consider the n x n matrix initial-value problem
12. (a) Find the solution of the initial-value problem
(b) Consider the difference equation
If
0
= 1 and z\ = 2, what is the value of Z IQ OO? What is the value of Zk in
general?
Exercises 123
(b) Solve the differential equation
i = Ax + b; x(O) = [ ~ l
9. Consider the initial-value problem
i(t) = Ax(t); x(O) = Xo
for t ~ O. Suppose that A E ~ n x n is skew-symmetric and let ex = Ilxol12. Show that
I/X(t)1/2 = ex for all t > O.
10. Consider the n x n matrix initial-value problem
X(t) = AX(t) - X(t)A; X(O) = c.
Show that the eigenvalues of the solution X (t) of this problem are the same as those
of C for all t.
11. The year is 2004 and there are three large "free trade zones" in the world: Asia (A),
Europe (E), and the Americas (R). Suppose certain multinational companies have
total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R. Each
year half of the Americas' money stays home, a quarter goes to Europe, and a quarter
goes to Asia. For Europe and Asia, half stays home and half goes to the Americas.
(a) Find the matrix M that gives
[
A] [A]
E =M E
R year k+1 R year k
(b) Find the eigenvalues and right eigenvectors of M.
(c) Find the distribution of the companies' assets at year k.
(d) Find the limiting distribution of the $40 trillion as the universe ends, i.e., as
k ---* +00 (i.e., around the time the Cubs win a World Series).
(Exercise adapted from Problem 5.3.11 in [24].)
12. (a) Find the solution of the initial-value problem
.Yet) + 2y(t) + yet) = 0; yeO) = 1, .YeO) = O.
(b) Consider the difference equation
Zk+2 + 2Zk+1 + Zk = O.
If Zo = 1 and ZI = 2, what is the value of ZIOOO? What is the value of Zk in
general?
Chapter 12
Generalized Eigenvalue
Problems
12.1 The Generalized Eigenvalue/Eigenvector Problem
In this chapter we consider the generalized eigenvalue problem
125
where A, B e C"
xn
. The standard eigenvalue problem considered in Chapter 9 obviously
corresponds to the special case that B = I.
Definition 12.1. A nonzero vector x e C" is a right generalized eigenvector of the pair
(A, B) with A, B e C
MX
" if there exists a scalar A. e C, called a generalized eigenvalue,
such that
Similarly, a nonzero vector y e C" is a left generalized eigenvector corresponding to an
eigenvalue X if
When the context is such that no confusion can arise, the adjective "generalized"
is usually dropped. As with the standard eigenvalue problem, if x [y] is a right [left]
eigenvector, then so is ax [ay] for any nonzero scalar a. e C.
Definition 12.2. The matrix A X B is called a matrix pencil (or pencil of the matrices A
and B).
As with the standard eigenvalue problem, eigenvalues for the generalized eigenvalue
problem occur where the matrix pencil A X B is singular.
Definition 12.3. The polynomial 7 r(A.) = det(A A.5) is called the characteristic poly-
nomial of the matrix pair (A, B) . The roots ofn(X .) are the eigenvalues of the associated
generalized eigenvalue problem.
Remark 12.4. When A, B e E"
xn
, the characteristic polynomial is obviously real, and
hence nonreal eigenvalues must occur in complex conjugate pairs.
Chapter 12
Generalized Eigenvalue
Problems
12.1 The Generalized Eigenvalue/Eigenvector Problem
In this chapter we consider the generalized eigenvalue problem
Ax = 'ABx,
where A, B E e
nxn
. The standard eigenvalue problem considered in Chapter 9 obviously
corresponds to the special case that B = I.
Definition 12.1. A nonzero vector x E en is a right generalized eigenvector of the pair
(A, B) with A, B E e
nxn
if there exists a scalar 'A E e, called a generalized eigenvalue,
such that
Ax = 'ABx. (12.1)
Similarly, a nonzero vector y E en is a left generalized eigenvector corresponding to an
eigenvalue 'A if
(12.2)
When the context is such that no confusion can arise, the adjective "generalized"
is usually dropped. As with the standard eigenvalue problem, if x [y] is a right [left]
eigenvector, then so is ax [ay] for any nonzero scalar a E <C.
Definition 12.2. The matrix A - 'AB is called a matrix pencil (or pencil of the matrices A
and B).
As with the standard eigenvalue problem, eigenvalues for the generalized eigenvalue
problem occur where the matrix pencil A - 'AB is singular.
Definition 12.3. The polynomial n('A) = det(A - 'AB) is called the characteristic poly-
nomial of the matrix pair (A, B). The roots ofn('A) are the eigenvalues of the associated
generalized eigenvalue problem.
Remark 12.4. When A, B E jRnxn, the characteristic polynomial is obviously real, and
hence nonreal eigenvalues must occur in complex conjugate pairs.
125
and there are again four cases to consider.
Case 1: a ^ 0, ft ^ 0. There are two eigenvalues, 1 and ^.
Case 2: a = 0, ft ^ 0. There is only one eigenvalue, 1 (of multiplicity 1).
Case 3: a ^ 0, f3 = 0. There are two eigenvalues, 1 and 0.
Case 4: a = 0, (3 = 0. All A 6 C are eigenvalues since det(B uA) = 0.
At least for the case of regular pencils, it is apparent where the "missing" eigenvalues have
gone in Cases 2 and 3. That is to say, there is a second eigenvalue "at infinity" for Case 3 of
A A.B, with its reciprocal eigenvalue being 0 in Case 3 of the reciprocal pencil B nA.
A similar reciprocal symmetry holds for Case 2.
While there are applications in system theory and control where singular pencils
appear, only the case of regular pencils is considered in the remainder of this chapter. Note
that A and/or B may still be singular. If B is singular, the pencil A KB always has
126 Chapter 12. Generalized Eigenvalue Problems
Remark 12.5. If B = I (or in general when B is nonsingular), then n ( X ) is a polynomial
of degree n, and hence there are n eigenvalues associated with the pencil A X B. However,
when B = I, in particular, when B is singular, there may be 0, k e n, or infinitely many
eigenvalues associated with the pencil A X B. For example, suppose
where a and ft are scalars. Then the characteristic polynomial is
and there are several cases to consider.
Case 1: a ^ 0, ft ^ 0. There are two eigenvalues, 1 and |.
Case 2: a = 0, f3 / 0. There are two eigenvalues, 1 and 0.
Case 3: a = 0, f3 = 0. There is only one eigenvalue, 1 (of multiplicity 1).
Case 4: a = 0, f3 = 0. All A e C are eigenvalues since det(A A. B ) =0.
Definition 12.6. If del (A X B) is not identically zero, the pencil A X B is said to be
regular; otherwise, it is said to be singular.
Note that if AA(A) n J\f(B) ^ 0, the associated matrix pencil is singular (as in Case
4 above).
Associated with any matrix pencil A X B is a reciprocal pencil B n,A and cor-
responding generalized eigenvalue problem. Clearly the reciprocal pencil has eigenvalues
(JL = . It is instructive to consider the reciprocal pencil associated with the example in
Remark 12.5. With A and B as in (12.3), the characteristic polynomial is
Remark 12.5. If B = I (or in general when B is nonsingular), then rr(A) is a polynomial
of degree n, and hence there are n eigenvalues associated with the pencil A - AB. However,
when B =I- I, in particular, when B is singular, there may be 0, k E !!, or infinitely many
eigenvalues associated with the pencil A - AB. For example, suppose
where a and (3 are scalars. Then the characteristic polynomial is
det(A - AB) = (I - AHa - (3A)
and there are several cases to consider.
Case 1: a =I- 0, {3 =I- O. There are two eigenvalues, I and ~ .
Case 2: a = 0, {3 =I- O. There are two eigenvalues, I and O.
Case 3: a =I- 0, {3 = O. There is only one eigenvalue, I (of multiplicity 1).
Case 4: a = 0, (3 = O. All A E C are eigenvalues since det(A - AB) == O.
(12.3)
Definition 12.6. If det(A - AB) is not identically zero, the pencil A - AB is said to be
regular; otherwise, it is said to be singular.
Note that if N(A) n N(B) =I- 0, the associated matrix pencil is singular (as in Case
4 above).
Associated with any matrix pencil A - AB is a reciprocal pencil B - /.LA and cor-
responding generalized eigenvalue problem. Clearly the reciprocal pencil has eigenvalues
/.L = . It is instructive to consider the reciprocal pencil associated with the example in
Remark 12.5. With A and B as in (12.3), the characteristic polynomial is
det(B - /.LA) = (1 - /.L)({3 - a/.L)
and there are again four cases to consider.
Case 1: a =I- 0, {3 =I- O. There are two eigenvalues, I and ~ .
Case 2: a = 0, {3 =I- O. There is only one eigenvalue, I (of multiplicity I).
Case 3: a =I- 0, {3 = O. There are two eigenvalues, 1 and O.
Case 4: a = 0, (3 = O. All A E C are eigenvalues since det(B - /.LA) == O.
At least for the case of regular pencils, it is apparent where the "missing" eigenvalues have
gone in Cases 2 and 3. That is to say, there is a second eigenvalue "at infinity" for Case 3 of
A - AB, with its reciprocal eigenvalue being 0 in Case 3 of the reciprocal pencil B - /.LA.
A similar reciprocal symmetry holds for Case 2.
While there are applications in system theory and control where singular pencils
appear, only the case of regular pencils is considered in the remainder of this chapter. Note
that A and/or B may still be singular. If B is singular, the pencil A - AB always has
12. 2. Canonical Forms 127
fewer than n eigenvalues. If B is nonsingular, the pencil A- A. f i always has precisely n
eigenvalues, since the generalized eigenvalue problem is then easily seen to be equivalent
to the standard eigenvalue problem B~
l
Ax = Xx (or AB~
l
w = Xw). However, this turns
out to be a very poor numerical procedure for handling the generalized eigenvalue problem
if B is even moderately ill conditioned with respect to inversion. Numerical methods that
work directly on A and B are discussed in standard textbooks on numerical linear algebra;
see, for example, [7, Sec. 7.7] or [25, Sec. 6.7].
12.2 Canonical Forms
Just as for the standard eigenvalue problem, canonical forms are available for the generalized
eigenvalue problem. Since the latter involves a pair of matrices, we now deal with equiva-
lencies rather than similarities, and the first theorem deals with what happens to eigenvalues
and eigenvectors under equivalence.
Theorem 12.7. Let A, fl, Q, Z e C
nxn
with Q and Z nonsingular. Then
1. the eigenvalues of the problems A XB and QAZ XQBZ are the same (the two
problems are said to be equivalent).
2. ifx is a right eigenvector of AXB, then Z~
l
x is a right eigenvector of QAZXQ B Z.
3. ify is a left eigenvector of A KB, then Q~
H
y isa left eigenvector ofQAZ XQBZ.
Proof:
1. det(QAZ-XQBZ) = det[0(A - XB)Z] = det gdet Zdet(A - XB). Since det 0
and det Z are nonzero, the result follows.
2. The result follows by noting that (A yB)x - Oif andonly if Q(A-XB)Z(Z~
l
x) =
0.
3. Again, the result follows easily by noting that y
H
(A XB) 0 if and only if
( Q~
H
y )
H
Q( A XB ) Z = Q. D
where T
a
and Tp are upper triangular.
By Theorem 12.7, the eigenvalues of the pencil A XB are then the ratios of the diag-
onal elements of T
a
to the corresponding diagonal elements of Tp, with the understanding
that a zero diagonal element of Tp corresponds to an infinite generalized eigenvalue.
There is also an analogue of the Murnaghan-Wintner Theorem for real matrices.
The first canonical form is an analogue of Schur's Theorem and forms, in fact, the
theoretical foundation for the QZ algorithm, which is the generally preferred method for
solving the generalized eigenvalue problem; see, for example, [7, Sec. 7.7] or [25, Sec. 6.7].
Theorem 12.8. Let A, B e Cn
xn
. Then there exist unitary matrices Q, Z e Cn
xn
such that
12.2. Canonical Forms 127
fewer than n eigenvalues. If B is nonsingular, the pencil A - AB always has precisely n
eigenvalues, since the generalized eigenvalue problem is then easily seen to be equivalent
to the standard eigenvalue problem B-
1
Ax = Ax (or AB-
1
W = AW). However, this turns
out to be a very poor numerical procedure for handling the generalized eigenvalue problem
if B is even moderately ill conditioned with respect to inversion. Numerical methods that
work directly on A and B are discussed in standard textbooks on numerical linear algebra;
see, for example, [7, Sec. 7.7] or [25, Sec. 6.7].
12.2 Canonical Forms
Just as for the standard eigenvalue problem, canonical forms are available for the generalized
eigenvalue problem. Since the latter involves a pair of matrices, we now deal with equiva-
lencies rather than similarities, and the first theorem deals with what happens to eigenvalues
and eigenvectors under equivalence.
Theorem 12.7. Let A, B, Q, Z E c
nxn
with Q and Z nonsingular. Then
1. the eigenvalues of the problems A - AB and QAZ - AQBZ are the same (the two
problems are said to be equivalent).
2. ifx isa right eigenvector of A-AB, then Z-l x isa righteigenvectorofQAZ-AQB Z.
3. ify isa left eigenvector of A -AB, then Q-H y isa lefteigenvectorofQAZ -AQBZ.
Proof:
1. det(QAZ - AQBZ) = det[Q(A - AB)Z] = det Q det Z det(A - AB). Since det Q
and det Z are nonzero, the result follows.
2. The result follows by noting that (A -AB)x = 0 if and only if Q(A -AB)Z(Z-l x) =
o.
3. Again, the result follows easily by noting that yH (A - AB) o if and only if
(Q-H y)H Q(A _ AB)Z = O. 0
The first canonical form is an analogue of Schur's Theorem and forms, in fact, the
theoretical foundation for the QZ algorithm, which is the generally preferred method for
solving the generalized eigenvalue problem; see, for example, [7, Sec. 7.7] or [25, Sec. 6.7].
Theorem 12.8. Let A, B E c
nxn
. Then there exist unitary matrices Q, Z E c
nxn
such that
QAZ = T
a
, QBZ = T
fJ
,
where Ta and TfJ are upper triangular.
By Theorem 12.7, the eigenvalues ofthe pencil A - AB are then the ratios of the diag-
onal elements of Ta to the corresponding diagonal elements of T
fJ
, with the understanding
that a zero diagonal element of TfJ corresponds to an infinite generalized eigenvalue.
There is also an analogue of the Murnaghan-Wintner Theorem for real matrices.
nxn
. Then there exist orthogonal matrices Q, Z e R"
xn
such
thnt
where T is upper triangular and S is quasi-upper-triangular.
When S has a 2 x 2 diagonal block, the 2 x 2 subpencil formed with the corresponding
2x2 diagonal subblock of T has a pair of complex conjugate eigenvalues. Otherwise, real
eigenvalues are given as above by the ratios of diagonal elements of S to corresponding
elements of T.
There is also an analogue of the Jordan canonical form called the Kronecker canonical
form (KCF). A full description of the KCF, including analogues of principal vectors and
so forth, is beyond the scope of this book. In this chapter, we present only statements of
the basic theorems and some examples. The first theorem pertains only to "square" regular
pencils, while the full KCF in all its generality applies also to "rectangular" and singular
pencils.
Theorem 12.10. Let A, B e C
nxn
and suppose the pencil A XB is regular. Then there
exist nonsingular matrices P, Q C"
x
" such that
where J is a Jordan canonical form corresponding to the finite eigenvalues of A -A.fi and
N is a nilpotent matrix of Jordan blocks associated with 0 and corresponding to the infinite
eigenvalues of A XB.
Example 12.11. The matrix pencil
with characteristic polynomial (X 2)
2
has a finite eigenvalue 2 of multiplicty 2 and three
infinite eigenvalues.
Theorem 12.12 (Kronecker Canonical Form). Let A, B e C
mxn
. Then there exist
nonsingular matrices P e C
mxm
and Q e C
nxn
such that
Theorem 12.9. Let A, B E jRnxn. Then there exist orthogonal matrices Q, Z E jRnxn such
that
QAZ = S, QBZ = T,
where T is upper triangular and S is quasi-upper-triangular.
When S has a 2 x 2 diagonal block, the 2 x 2 subpencil fonned with the corresponding
2 x 2 diagonal subblock of T has a pair of complex conjugate eigenvalues. Otherwise, real
eigenvalues are given as above by the ratios of diagonal elements of S to corresponding
elements of T.
There is also an analogue of the Jordan canonical fonn called the Kronecker canonical
form (KeF). A full description of the KeF, including analogues of principal vectors and
so forth, is beyond the scope of this book. In this chapter, we present only statements of
the basic theorems and some examples. The first theorem pertains only to "square" regular
pencils, while the full KeF in all its generality applies also to "rectangular" and singular
pencils.
Theorem 12.10. Let A, B E c
nxn
and suppose the pencil A - AB is regular. Then there
exist nonsingular matrices P, Q E c
nxn
such that
peA - AB)Q = [ ~ ~ ] - A ~ ~ l
where J is a Jordan canonical form corresponding to the finite eigenvalues of A - AB and
N is a nilpotent matrix of Jordan blocks associated with 0 and corresponding to the infinite
eigenvalues of A - AB.
Example 12.11. The matrix pencil
[2 I
0 0
~ ]-> [
0 0
o 0] o 2 0 0 I 0 o 0
o 0 1 0 0 0 I 0
o 0 0 1 0 0 o 0
o 0 0 0 0 0 0 0
with characteristic polynomial (A - 2)2 has a finite eigenvalue 2 of multiplicty 2 and three
infinite eigenvalues.
Theorem 12.12 (Kronecker Canonical Form). Let A, B E c
mxn
Then there exist
nonsingular matrices P E c
mxm
and Q E c
nxn
such that
peA - AB)Q = diag(LII' ... , L
l
" L ~ , ... L;'. J - A.I, I - )"N),
where N is nilpotent, both N and J are in Jordan canonical form, and L^ is the (k + 1) x k
bidiagonal pencil
The /( are called the left minimal indices while the r, are called the right minimal indices.
Left or right minimal indices can take the value 0.
Such a matrix is in KCF. The first block of zeros actually corresponds to LQ, LQ, LQ, LQ ,
LQ, where each LQ has "zero columns" and one row, while each LQ has "zero rows" and
one column. The second block is L\ while the third block is L\. The next two blocks
correspond to
Just as sets of eigenvectors span A-invariant subspaces in the case of the standard
eigenproblem (recall Definition 9.35), there is an analogous geometric concept for the
generalized eigenproblem.
Definition 12.14. Let A, B e W
lxn
and suppose the pencil A XB is regular. Then V is a
deflating subspace if
Just as in the standard eigenvalue case, there is a matrix characterization of deflating
subspace. Specifically, suppose S e R n*
xk
is a matrix whose columns span a ^-dimensional
subspace S of R
n
, i.e., R ( S) = <S. Then S is a deflating subspace for the pencil A XB if
and only if there exists M e R
kxk
such that
while the nilpotent matrix N in this example is
where N is nilpotent, both Nand J are in Jordan canonical form, and Lk is the (k + I) x k
bidiagonal pencil
-A 0 0
-A
Lk =
0 0
-A
0 0 I
The Ii are called the left minimal indices while the ri are called the right minimal indices.
Left or right minimal indices can take the value O.
Example 12.13. Consider a 13 x 12 block diagonal matrix whose diagonal blocks are
-A 0]
I -A .
o I
Such a matrix is in KCF. The first block of zeros actually corresponds to Lo, Lo, Lo, L6,
L6, where each Lo has "zero columns" and one row, while each L6 has "zero rows" and
one column. The second block is L\ while the third block is LI- The next two blocks
correspond to
[
21
J = 0 2
o 0
while the nilpotent matrix N in this example is

000
Just as sets of eigenvectors span A-invariant subspaces in the case of the standard
eigenproblem (recall Definition 9.35), there is an analogous geometric concept for the
generalized eigenproblem.
Definition 12.14. Let A, B E and suppose the pencil A - AB is regular. Then V is a
deflating subspace if
dim(AV + BV) = dimV. (12.4)
Just as in the standard eigenvalue case, there is a matrix characterization of deflating
subspace. Specifically, suppose S E is a matrix whose columns span a k-dimensional
subspace S of i.e., n(S) = S. Then S is a deflating subspace for the pencil A - AB if
and only if there exists M E such that
AS = BSM. (12.5)
If B = /, then (12.4) becomes dim(AV + V) = dimV, which is clearly equivalent to
AV c V. Similarly, (12.5) becomes AS = SM as before. If the pencil is not regular, there
is a concept analogous to deflating subspace called a reducing subspace.
12.3 Application to the Computation of System Zeros
Consider the linear svstem
which has a root at 2.8 .
The method of finding system zeros via a generalized eigenvalue problem also works
well for general multi-input, multi-output systems. Numerically, however, one must be
careful first to "deflate out" the infinite zeros (infinite eigenvalues of (12.6)). This is accom-
plished by computing a certain unitary equivalence on the system pencil that then yields a
smaller generalized eigenvalue problem with only finite generalized eigenvalues (the finite
zeros).
The connection between system zeros and the corresponding system pencil is non-
trivial. However, we offer some insight below into the special case of a single-input,
with A M
n x n
, B R"
x m
, C e R
pxn
, and D R
pxm
. This linear time-invariant state-
space model is often used in multivariable control theory, where x(= x(t)) is called the state
vector, u is the vector of inputs or controls, and y is the vector of outputs or observables.
For details, see, for example, [26].
In general, the (finite) zeros of this system are given by the (finite) complex numbers
z, where the "system pencil"
drops rank. In the special case p = m, these values are the generalized eigenvalues of the
(n + m) x (n + m) pencil.
Example 12.15. Let
Then the transfer matrix (see [26]) of this system is
which clearly has a zero at 2.8 . Checking the finite eigenvalues of the pencil (12.6), we
find the characteristic polynomial to be
If B = I, then (12.4) becomes dim (A V + V) = dim V, which is clearly equivalent to
AV ~ V. Similarly, (12.5) becomes AS = SM as before. lEthe pencil is not regular, there
is a concept analogous to deflating subspace called a reducing subspace.
12.3 Application to the Computation of System Zeros
Consider the linear system
i = Ax + Bu,
y = Cx + Du
with A E jRnxn, B E jRnxm, C E jRPxn, and D E jRPxm. This linear time-invariant state-
space model is often used in multivariable control theory, where x(= x(t)) is called the state
vector, u is the vector of inputs or controls, and y is the vector of outputs or observables.
For details, see, for example, [26].
In general, the (finite) zeros of this system are given by the (finite) complex numbers
z, where the "system pencil"
(12.6)
drops rank. In the special case p = m, these values are the generalized eigenvalues of the
(n + m) x (n + m) pencil.
Example 12.15. Let
A=[
-4
2
Then the transfer matrix (see [26)) of this system is
C = [I 2],
55 + 14
g(5)=C(sI-A)-'B+D= 2 '
5 + 3s + 2
D=O.
which clearly has a zero at -2.8. Checking the finite eigenvalues of the pencil (12.6), we
find the characteristic polynomial to be
det
[
A -c
M
B]
D "'" 5A + 14,
which has a root at -2.8.
The method of finding system zeros via a generalized eigenvalue problem also works
well for general mUlti-input, multi-output systems. Numerically, however, one must be
careful first to "deflate out" the infinite zeros (infinite eigenvalues of (12.6. This is accom-
plished by computing a certain unitary equivalence on the system pencil that then yields a
smaller generalized eigenvalue problem with only finite generalized eigenvalues (the finite
zeros).
The connection between system zeros and the corresponding system pencil is non-
trivial. However, we offer some insight below into the special case of a single-input.
12.4. Symmetric Generalized Eigenvalue Problems 131
single-output system. Specifically, let B = b e Rn, C = c
1
e R
l xn
, and D = d e R.
Furthermore, let g(.s) = c
r
(s7 A )~
!
Z ? + d denote the system transfer function (matrix),
and assume that g ( s ) can be written in the form
where T T (S ) is the characteristic polynomial of A, and v(s) and T T (S ) are relatively prime
(i.e., there are no "pole/zero cancellations").
Suppose z C is such that
is singular. Then there exists a nonzero solution to
or
Assuming z is not an eigenvalue of A (i.e., no pole/zero cancellations), then from (12.7) we
get
Substituting this in (12.8), we have
or g ( z ) y = 0 by the definition of g . Now _ y ^ 0 (else x = 0 from (12.9)). Hence g(z) = 0,
i.e., z is a zero of g.
12.4 Symmetric Generalized Eigenvalue Problems
A very important special case of the generalized eigenvalue problem
for A, B e R
nxn
arises when A = A and B = B
1
> 0. For example, the second-order
system of differential equations
where M is a symmetric positive definite "mass matrix" and K is a symmetric "stiffness
matrix," is a frequently employed model of structures or vibrating systems and yields a
generalized eigenvalue problem of the form (12.10).
Since B is positive definite it is nonsingular. Thus, the problem (12.10) is equivalent
to the standard eigenvalue problem B~
l
Ax = A J C. However, B~
1
A is not necessarily
symmetric.
12.4. Symmetric Generalized Eigenvalue Problems 131
single-output system. Specifically, let B = b E ffi.n, C = c
T
E ffi.l xn, and D = d E R
Furthermore, let g(s) = c
T
(s I - A) -1 b + d denote the system transfer function (matrix),
and assume that g(s) can be written in the form
v(s)
g(s) = n(s)'
where n(s) is the characteristic polynomial of A, and v(s) and n(s) are relatively prime
(i.e., there are no "pole/zero cancellations").
Suppose Z E C is such that
[
A - zI b ]
c
T
d
is singular. Then there exists a nonzero solution to
or
(A - zl)x + by = 0,
c
T
x +dy = O.
(12.7)
(12.8)
Assuming z is not an eigenvalue of A (i.e., no pole/zero cancellations), then from (12.7) we
get
x = -(A - zl)-lby.
(12.9)
Substituting this in (12.8), we have
_c
T
(A - zl)-lby + dy = 0,
or g(z)y = 0 by the definition of g. Now y 1= 0 (else x = 0 from (12.9. Hence g(z) = 0,
i.e., z is a zero of g.
12.4 Symmetric Generalized Eigenvalue Problems
A very important special case of the generalized eigenvalue problem
Ax = ABx (12.10)
for A, B E ffi.nxn arises when A = AT and B = BT > O. For example, the second-order
system of differential equations
Mx+Kx=O,
where M is a symmetric positive definite "mass matrix" and K is a symmetric "stiffness
matrix," is a frequently employed model of structures or vibrating systems and yields a
generalized eigenvalue problem ofthe form (12.10).
Since B is positive definite it is nonsingular. Thus, the problem (12.10) is equivalent
to the standard eigenvalue problem B-
1
Ax = AX. However, B-
1
A is not necessarily
symmetric.
Nevertheless, the eigenvalues of B
l
A are always real (and are approximately 2.1926
and -3.1926 in Example 12.16).
nxn
with A = A
T
and B = B
T
> 0. Then the generalized
eigenvalue problem
whose eigenvalues are approximately 2.1926 and 3.1926 as expected.
The material of this section can, of course, be generalized easily to the case where A
and B are Hermitian, but since real-valued matrices are commonly used in most applications,
we have restricted our attention to that case only.
has n real eigenvalues, and the n corresponding right eigenvectors can be chosen to be
orthogonal with respect to the inner product (x, y)
B
= X
T
By. Moreover, if A > 0, then
the eigenvalues are also all positive.
Proof: Since B > 0, it has a Cholesky factorization B = LL
T
, where L is nonsingular
(Theorem 10.23). Then the eigenvalue problem
can be rewritten as the equivalent problem
Letting C = L
1
AL
J
and z = L
1
x, (12.11) can then be rewritten as
Since C = C
T
, the eigenproblem (12.12) has n real eigenvalues, with corresponding eigen-
vectors zi,..., z
n
satisfying
Then x, = L
T
zi, i n, are eigenvectors of the original generalized eigenvalue problem
and satisfy
Finally, if A = A
T
> 0, then C = C
T
> 0, so the eigenvalues are positive. D
Example 12.18. The Cholesky factor for the matrix B in Example 12.16 is
Then it is easily checked thai
Example 12.16. Let A ThenB~
l
A
Example 12.16. Let A = ; l B = [i J Then A = J
Nevertheless, the eigenvalues of A are always real (and are approximately 2.1926
and -3.1926 in Example 12.16).
Theorem 12.17. Let A, B E jRnxn with A = AT and B = BT > O. Then the generalized
eigenvalue problem
Ax = ABx
has n real eigenvalues, and the n corresponding right eigenvectors can be chosen to be
orthogonal with respect to the inner product (x, y) B = x
T
By. Moreover, if A > 0, then
the eigenvalues are also all positive.
Proof: Since B > 0, it has a Cholesky factorization B = LL T, where L is nonsingular
(Theorem 10.23). Then the eigenvalue problem
Ax = ABx = ALL T x
can be rewritten as the equivalent problem
(12.11)
Letting C = L AL and Z = LT x, (12.11) can then be rewritten as
Cz = AZ. (12.12)
Since C = C
T
, the eigenproblem (12.12) has n real eigenvalues, with corresponding eigen-
vectors Z I, .. , Zn satisfying
zi Zj = Dij.
Then Xi = L Zi, i E !!., are eigenvectors of the original generalized eigenvalue problem
and satisfy
(Xi, Xj)B = xr BXj = (zi L Zj) = Dij.
Finally, if A = AT> 0, then C = C
T
> 0, so the eigenvalues are positive. 0
Example 12.18. The Cholesky factor for the matrix B in Example 12.16 is
1] .
.,fi .,fi
Then it is easily checked that
c = = [ 0 . .5
2.5
2 . .5 ]
-1.5 '
whose eigenvalues are approximately 2.1926 and -3.1926 as expected.
The material of this section can, of course, be generalized easily to the case where A
and B are Hermitian, but since real-valued matrices are commonly used in most applications,
we have restricted our attention to that case only.
12.5. Simultaneous Diagonalization 133
12.5 Simultaneous Diagonalization
Recall that many matrices can be diagonalized by a similarity. In particular, normal ma-
trices can be diagonalized by a unitary similarity. It turns out that in some cases a pair of
matrices (A, B) can be simultaneously diagonalized by the same matrix. There are many
such results and we present only a representative (but important and useful) theorem here.
Again, we restrict our attention only to the real case, with the complex case following in a
straightforward way.
Theorem 12.19 (Simultaneous Reduction to Diagonal Form). Let A, B e E"
x
" with
A = A
T
and B = B
T
> 0. Then there exists a nonsingular matrix Q such that
\ 2.5.1 Simultaneous diagonalization via SVD
There are situations in which forming C = L~
1
AL~
T
as in the proof of Theorem 12.19 is
numerically problematic, e.g., when L is highly ill conditioned with respect to inversion. In
such cases, simultaneous reduction can also be accomplished via an SVD. To illustrate, let
where D is diagonal. In fact, the diagonal elements of D are the eigenvalues of B
1
A.
Proof: Let B = LL
T
be the Cholesky factorization of B and setC = L~
1
AL~
T
. Since
C is symmetric, there exists an orthogonal matrix P such that P
T
CP = D, where D is
diagonal. Let Q = L~
T
P. Then
and
Finally, since QDQ~
l
= QQ
T
AQQ~
l
= L-
T
PP
T
L~
1
A = L~
T
L~
1
A = B~
1
A, we
haveA(D) = A(B~
1
A). D
Note that Q is not in general orthogonal, so it does not preserve eigenvalues of A and B
individually. However, it does preserve the eigenvalues of A XB. This can be seen directly.
LetA = Q
T
AQandB = Q
T
BQ. Then/HA = Q~
l
B~
l
Q~
T
Q
T
AQ = Q~
1
B~
1
AQ.
Theorem 12.19 is very useful for reducing many statements about pairs of symmetric
matrices to "the diagonal case." The following is typical.
Theorem 12.20. Let A, B e M"
xn
be positive definite. Then A > B if and only if B~
l
>
A-
1
.
Proof: By Theorem 12.19, there exists Q e E"
x
" such that Q
T
AQ = D and Q
T
BQ = I,
where D is diagonal. Now D > 0 by Theorem 10.31. Also, since A > B, by Theorem
10.21 we have that Q
T
AQ > Q
T
BQ, i.e., D > I. But then D"
1
< / (this is trivially true
since the two matrices are diagonal). Thus, QD~
l
Q
T
< QQ
T
, i.e., A~
l
< B~
l
. D
12.5. Simultaneous Diagonalization 133
12.5 Simultaneous Diagonalization
Recall that many matrices can be diagonalized by a similarity. In particular, normal ma-
trices can be diagonalized by a unitary similarity. It turns out that in some cases a pair of
matrices (A, B) can be simultaneously diagonalized by the same matrix. There are many
such results and we present only a representative (but important and useful) theorem here.
Again, we restrict our attention only to the real case, with the complex case following in a
straightforward way.
Theorem 12.19 (Simultaneous Reduction to Diagonal Form). Let A, B E ] [ ~ n x n with
A = AT and B = BT > O. Then there exists a nonsingular matrix Q such that
where D is diagonal. Infact, the diagonal elements of D are the eigenvalues of B-
1
A.
Proof: Let B = LLT be the Cholesky factorization of B and set C = L -I AL -T. Since
C is symmetric, there exists an orthogonal matrix P such that pTe p = D, where D is
diagonal. Let Q = L - T P. Then
and
QT BQ = pT L -I(LLT)L -T P = pT P = [.
Finally, since QDQ-I = QQT AQQ-I = L -T P pT L -I A = L -T L -I A
B-
1
A, we
have A(D) = A(B-
1
A). 0
Note that Q is not in general orthogonal, so it does not preserve eigenvalues of A and B
individually. However, it does preserve the eigenvalues of A - 'AB. This can be seen directly.
Let A = QT AQ and B = QT BQ. Then B-
1
A = Q-1 B-
1
Q-T QT AQ = Q-1 B-
1
AQ.
Theorem 12.19 is very useful for reducing many statements about pairs of symmetric
matrices to "the diagonal case." The following is typical.
Theorem 12.20. Let A, B E lR
nxn
be positive definite. Then A 2: B if and only if B-
1
2:
A-I.
Proof: By Theorem 12.19, there exists Q E l R ~ x n such that QT AQ = D and QT BQ = [,
where D is diagonal. Now D > 0 by Theorem 10.31. Also, since A 2: B, by Theorem
10.21 we have that QT AQ 2: QT BQ, i.e., D 2: [. But then D-
I
:::: [(this is trivially true
since the two matrices are diagonal). Thus, Q D-
I
QT :::: Q QT, i.e., A -I :::: B-
1
. 0
12.5.1 Simultaneous diagonalization via SVD
There are situations in which forming C = L -I AL -T as in the proof of Theorem 12.19 is
numerically problematic, e.g., when L is highly iII conditioned with respect to inversion. In
such cases, simultaneous reduction can also be accomplished via an SVD. To illustrate. let
The problem (12.15) is called a generalized singular value problem and algorithms exist to
solve it (and hence equivalently (12.13)) via arithmetic operations performed only on LA
and LB separately, i.e., without forming the products L
A
L
T
A
or L
B
L
T
B
explicitly; see, for
example, [7, Sec. 8.7.3]. This is analogous to finding the singular values of a matrix M by
operations performed directly on M rather than by forming the matrix M
T
M and solving
the eigenproblem M
T
MX = Xx.
Remark 12.22. Various generalizations of the results in Remark 12.21 are possible, for
example, when A = A
T
> 0. The case when A is symmetric but indefinite is not so
straightforward, at least in real arithmetic. For example, A can be written as A = PDP
T
,
~ ~ ~ ~ T
where D is diagonal and P is orthogonal, but in writing A PDDP = PD(PD) with
D diagonal, D may have pure imaginary elements.
us assume that both A and B are positive definite. Further, let A = L
A
L
T
A
and B LsL
T
B
be Cholesky factorizations of A and B, respectively. Compute the SVD
where E e R
x
" is diagonal. Then the matrix Q = L
B
T
U performs the simultaneous
diagonalization. To check this, note that
while
Remark 12.21. The SVD in (12.13) can be computed without explicitly forming the
indicated matrix product or the inverse by using the so-called generalized singular value
decomposition (GSVD). Note that
and thus the singular values of L
B
L
A
can be found from the eigenvalue problem
Letting x = L
B
z we see that (12.14) can be rewritten in the form L
A
L
A
x = XL
B
z =
ALgL^Lg
7
z, which is thus equivalent to the generalized eigenvalue problem
us assume that both A and B are positive definite. Further, let A = and B =
be Cholesky factorizations of A and B, respectively. Compute the SVD
(12.13)
where L E xn is diagonal. Then the matrix Q = L i/ u performs the simultaneous
diagonalization. To check this, note that
while
QT AQ = U
T

= UTULVTVLTUTU
= L2
QT BQ = U
T

= UTU
= I.
Remark 12.21. The SVD in (12.13) can be computed without explicitly forming the
indicated matrix product or the inverse by using the so-called generalized singular value
decomposition (GSVD). Note that
and thus the singular values of L B 1 L A can be found from the eigenvalue problem
02.14)
Letting x = LBT Z we see that 02.14) can be rewritten in the form = ALBz =
z, which is thus equivalent to the generalized eigenvalue problem
02.15)
The problem (12.15) is called a generalized singular value problem and algorithms exist to
solve it (and hence equivalently (12.13 via arithmetic operations performed only on LA
and L B separately, i.e., without forming the products LA L or L B L explicitly; see, for
example, [7, Sec. 8.7.3]. This is analogous to finding the singular values of a matrix M by
operations performed directly on M rather than by forming the matrix MT M and solving
the eigenproblem MT M x = AX.
Remark 12.22. Various generalizations of the results in Remark 12.21 are possible, for
example, when A = AT::: O. The case when A is symmetric but indefinite is not so
straightforward, at least in real arithmetic. For example, A can be written as A = PDP T,
where Disdiagonaland P is orthogonal,butin writing A = PDDp
T
= PD(PD{ with
D diagonal, b may have pure imaginary elements.
12.6. Higher-Order Eigenvalue Problems 135
12.6 Higher-Order Eigenvalue Problems
Consider the second-order system of differential equations
where q(t} e W
1
and M, C, K e Rn
xn
. Assume for simplicity that M is nonsingular.
Suppose, by analogy with the first-order case, that we try to find a solution of (12.16) of the
form q(t) = e
xt
p, where the n-vector p and scalar A. are to be determined. Substituting in
(12.16) we get
To get a nonzero solution /?, we thus seek values of A. for which the matrix A.
2
M + A.C + K
is singular. Since the determinantal equation
yields a polynomial of degree 2rc, there are 2n eigenvalues for the second-order (or
quadratic) eigenvalue problem A.
2
M + A.C + K.
A special case of (12.16) arises frequently in applications: M = I, C = 0, and
K = K
T
. Suppose K has eigenvalues
Let a > k = | f j i k 1
2
Then the 2n eigenvalues of the second-order eigenvalue problem A.
2
/ + K
are
If r = n (i.e., K = K
T
> 0), then all solutions of q + Kq = 0 are oscillatory.
Let x\ = q and \i = q. Then (12.16) can be written as a first-order system (with block
companion matrix)
where x(t) . E
2
". If M is singular, or if it is desired to avoid the calculation of M
l
because
M is too ill conditioned with respect to inversion, the second-order problem (12.16) can still
be converted to the first-order generalized linear system
or, since
12.6. Higher-Order Eigenvalue Problems 135
12.6 Higher-Order Eigenvalue Problems
Consider the second-order system of differential equations
Mq+Cq+Kq=O, (12.16)
where q(t) E ~ n and M, C, K E ~ n x n . Assume for simplicity that M is nonsingular.
Suppose, by analogy with the first-order case, that we try to find a solution of (12.16) of the
form q(t) = eAt p, where the n-vector p and scalar A are to be determined. Substituting in
(12.16) we get
or, since eAt :F 0,
(A
2
M + AC + K) p = O.
To get a nonzero solution p, we thus seek values of A for which the matrix A
2
M + AC + K
is singular. Since the determinantal equation
o = det(A
2
M + AC + K) = A 2n + ...
yields a polynomial of degree 2n, there are 2n eigenvalues for the second-order (or
quadratic) eigenvalue problem A
2
M + AC + K.
A special case of (12.16) arises frequently in applications: M = I, C = 0, and
K = KT. Suppose K has eigenvalues
IL I ::: ... ::: ILr ::: 0 > ILr+ I ::: ... ::: ILn
Let Wk = I ILk I !. Then the 2n eigenvalues of the second-order eigenvalue problem A
2
I + K
are
jWk; k = 1, ... , r,
Wk; k = r + 1, ... , n.
If r = n (i.e., K = KT ::: 0), then all solutions of q + K q = 0 are oscillatory.
Let XI = q and X2 = q. Then (12.16) can be written as a first-order system (with block
companion matrix)
. [ 0
X = -M-1K
where x (t) E ~ 2 n . If M is singular, or if it is desired to avoid the calculation of M-
I
because
M is too ill conditioned with respect to inversion, the second-order problem (12.16) can still
be converted to the first-order generalized linear system
[
I OJ' [0 I J
o M x = -K -C x.
Many other first-order realizations are possible. Some can be useful when M, C, and/or K
have special symmetry or skew-symmetry properties that can exploited.
Higher-order analogues of (12.16) involving, say, the kth derivative of q, lead naturally
to higher-order eigenvalue problems that can be converted to first-order form using aknxkn
block companion matrix analogue of (11.19). Similar procedures hold for the general k\h-
order difference equation
EXERCISES
are the eigenvalues of the matrix A BD
1
C.
2. Let F, G C
MX
". Show that the nonzero eigenvalues of FG and GF are the same.
Hint: An easy "trick proof is to verify that the matrices
are similar via the similarity transformation
are identical for all F 6 E"
1
*" and all G G R"
xm
.
Hint: Consider the equivalence
(A similar result is also true for "nonsquare" pencils. In the parlance of control theory,
such results show that zeros are invariant under state feedback or output injection.)
which can be converted to various first-order systems of dimension kn.
1. Suppose A e R
nx
" and D e R
xm
. Show that the finite generalized eigenvalues of
the pencil
3. Let F e C
nxm
, G e C
mx
". Are the nonzero singular values of FG and GF the
same?
4. Suppose A R
nxn
, B e R
n
*
m
, and C e E
wx
". Show that the generalized eigenval-
ues of the pencils
and
Many other first-order realizations are possible. Some can be useful when M, C, andlor K
have special symmetry or skew-symmetry properties that can exploited.
Higher-order analogues of (12.16) involving, say, the kth derivative of q, lead naturally
to higher-order eigenvalue problems that can be converted to first-order form using a kn x kn
block companion matrix analogue of (11.19). Similar procedures hold for the general kth-
order difference equation
which can be converted to various first-order systems of dimension kn.
EXERCISES
1. Suppose A E lR
n
xn and D E lR::! xm. Show that the finite generalized eigenvalues of
the pencil
[ ~ ~ J - A [ ~ ~ J
are the eigenvalues of the matrix A - B D-
1
C.
2. Let F, G E e
nxn
Show that the nonzero eigenvalues of FG and G F are the same.
Hint: An easy "trick proof' is to verify that the matrices
[Fg ~ ] and [ ~ GOF ]
are similar via the similarity transformation
3. Let F E e
nxm
, G E e
mxn
Are the nonzero singular values of FG and GF the
same?
4. Suppose A E ]Rnxn, B E lR
nxm
, and C E lRmxn. Show that the generalized eigenval-
ues of the pencils
[ ~ ~ J - A [ ~ ~ J
and
[ A + B ~ + GC ~ ] _ A ~ ~ ]
are identical for all F E Rm xn and all G E R" xm .
Hint: Consider the equivalence
[
I G][A-U B][I 0]
01 CO Fl'
(A similar result is also true for "nonsquare" pencils. In the parlance of control theory,
such results show that zeros are invariant under state feedback or output injection.)
Exercises 137
5. Another family of simultaneous diagonalization problems arises when it is desired
that the simultaneous diagonalizing transformation Q operates on matrices A, B e
]R
nx
" in such a way that Q~
l
AQ~
T
and Q
T
BQ are simultaneously diagonal. Such
a transformation is called contragredient. Consider the case where both A and
B are positive definite with Cholesky factorizations A = L&L
T
A
and B = L#Lg,
respectively, and let UW
T
be an SVD of L
T
B
L
A
.
(a) Show that Q = LA V~
5
is a contragredient transformation that reduces both
A and B to the same diagonal matrix.
(b) Show that Q~
l
= ^~Û
T
L
T
B
.
(c) Show that the eigenvalues of A B are the same as those of E
2
and hence are
positive.
Exercises 137
5. Another family of simultaneous diagonalization problems arises when it is desired
that the simultaneous diagonalizing transformation Q operates on matrices A, B E
jRnxn in such a way that Q-l AQ-T and QT BQ are simultaneously diagonal. Such
a transformation is called contragredient. Consider the case where both A and
B are positive definite with Cholesky factorizations A = LA L and B = L B L
respectively, and let be an SVD of
(a) Show that Q = LA is a contragredient transformation that reduces both
A and B to the same diagonal matrix.
(b) Show that Q-l =
(c) Show that the eigenvalues of AB are the same as those of 1;2 and hence are
positive.
Chapter 13
Kronecker Products
mx
", B e R
pxq
. Then the Kronecker product (or tensor
product) of A and B is defined as the matrix
Obviously, the same definition holds if A and B are complex-valued matrices. We
restrict our attention in this chapter primarily to real-valued matrices, pointing out the
extension to the complex case only where it is not obvious.
Example 13.2.
Note that B < g> A / A < g> B.
2. Foranyfl e!F
X(
7, /
2
< 8 > f l = [o l\
Replacing I
2
by / yields a block diagonal matrix with n copies of B along the
diagonal.
3. Let B be an arbitrary 2x2 matrix. Then
139
Chapter 13
Kronecker Products
Definition 13.1. Let A E lR
mxn
, B E lR
pxq
. Then the Kronecker product (or tensor
product) of A and B is defined as the matrix
[
allB
A@B= :
amlB
alnB ]
: E lRmpxnq.
amnB
(13.1)
Obviously, the same definition holds if A and B are complex-valued matrices. We
restrict our attention in this chapter primarily to real-valued matrices, pointing out the
extension to the complex case only where it is not obvious.
Example 13.2.
1. Let A =
2
nand B = [; Then
2

4 2 6
n
A@B = [
2B 3 4 6 6
2B 3 4 2 2
9 4 6 2
Note that B @ A i- A @ B.
2. Forany B E lR
pxq
, /z @ B = J.
Replacing 12 by In yields a block diagonal matrix with n copies of B along the
diagonal.
3. Let B be an arbitrary 2 x 2 matrix. Then
l b"
0
b12
0
l
B @/z =
b
ll
0 b12
0
b
2
2
0
b
21
0 b
22
139
140 Chapter 13. Kronecker Products
The extension to arbitrary B and / is obvious.
4. Let Jt R
m
, y e R". Then
5. Let* eR
m
, y eR". Then
13.2 Properties of the Kronecker Product
mx
", 5 e R
rxi
, C e R"
x
^ and D e R
sxt
. Then
Proof: Simply verify that
Theorem 13.4. For all A and B,
Proof: For the proof, simply verify using the definitions of transpose and Kronecker
product. D
Corollary 13.5. If A e R"
xn
and B e R
mxm
are symmetric, then A B is symmetric.
Theorem 13.6. If A and B are nonsingular,
Proof: Using Theorem 13.3, simply note that
The extension to arbitrary B and In is obvious.
4. Let x E y E !R.n. Then
[
T T]T
X Y = XIY , ... , XmY
= [XIYJ, ... , XIYn, X2Yl, ... , xmYnf E !R.
mn
.
13.2 Properties of the Kronecker Product
Theorem 13.3. Let A E B E C E and D E Then
(A 0 B)(C 0 D) = AC 0 BD (E
Proof; Simply verify that

=AC0BD. 0
Theorem 13.4. Foral! A and B, (A Bl = AT BT.
al;kCkPBD ]
amkckpBD
(13.2)
Proof' For the proof, simply verify using the definitions of transpose and Kronecker
product. 0
Corollary 13.5. If A E ]Rn xn and B E !R.
m
xm are symmetric, then A B is symmetric.
Theorem 13.6. If A and Bare nonsingular, (A B)-I = A-I B-
1
.
Proof: Using Theorem 13.3, simply note that (A B)(A -1 B-
1
) = 1 1 = I. 0
Corollary 13.8. If A E"
xn
is orthogonal and B e M
mxm
15 orthogonal, then A < g > B is
orthogonal.
13.2. Properties of the Kronecker Product 141
Theorem 13.7. If A e IR"
xn
am/ B eR
mxm
are normal, then A B is normal.
Proof:
yields a singular value decomposition of A < 8 > B (after a simple reordering of the diagonal
elements O/A < 8 > 5 and the corresponding right and left singular vectors).
Corollary 13.11. Let A e R
x
" have singular values a\ > > a
r
> 0 and let B e
have singular values T\ > > T
S
> 0. Then A < g ) B (or B < 8 > A) has rs singular values
îT\ > > ff
r
T
s
> Qand
nx
" have eigenvalues A., - , / e n, and let B e R
mxw
/zave
eigenvalues jJij, 7 m. TTzen ?/ze mn eigenvalues of A B are
Moreover, if x\, ..., x
p
are linearly independent right eigenvectors of A corresponding
to A - i , . . . , A.
p
(p < n), and zi, , z
q
are linearly independent right eigenvectors of B
corresponding to JJL\ , ..., \Ju
q
(q < m), then ;c, < 8 > Zj ffi.
m
" are linearly independent right
eigenvectors of A B corresponding to A., /u ,
7
, i e /?, 7 e q.
Proof: The basic idea of the proof is as follows:
If A and B are diag onalizable in Theorem 13.12, we can take p = n and q mand
thu s g et the complete eig enstru ctu re of A < 8 > B. In g eneral, if A and fi have Jordan form
Example 13.9. Let A and B - Then it is easily seen that
A i s orthog onal wi th eig envalu es e
j9
and B i s orthog onal wi th eig envalu es e
j(i>
. T he 4x4
matrix A 5 is then also orthog onal with eig envalu es e^'^+'W and e
^
( 6>
~^
>
\
Theorem 13.10. Lg f A G E
mx
" have a singular value decomposition l/Ê^Vj an^ /ef
fi e ^
pxq
have a singular value decomposition UB^B^B- Then
13.2. Properties of the Kronecker Product
Theorem 13.7. If A E IR
nxn
and B E IR
mxm
are normal, then A 0 B is normal.
Proof:
(A 0 B{ (A 0 B) = (AT 0 BT)(A 0 B) by Theorem 13.4
= AT A 0 BT B by Theorem 13.3
= AAT 0 B BT since A and B are normal
= (A 0 B)(A 0 B)T by Theorem 13.3. 0
141
Corollary 13.8. If A E IR
nxn
is orthogonal and B E IR
mxm
is orthogonal, then A 0 B is
orthogonal.
E I 139 L A
[
eose Sine] dB [Cos</> Sin</>] Th ., '1 h
xamp e . et = _ sin e cose an = _ sin</> cos</>O en It IS easl y seen t at
A is orthogonal with eigenvalues ejO and B is orthogonal with eigenvalues ej</J. The 4 x 4
matrix A 0 B is then also orthogonal with eigenvalues ejeH</ and eje
fJ
-</.
mxn
have a singular value decomposition VA ~ A vI and let
B E IR
pxq
have a singular value decomposition V B ~ B VI. Then
yields a singular value decomposition of A 0 B (after a simple reordering of the diagonal
elements of A 0 ~ B and the corresponding right and left singular vectors).
Corollary 13.11. Let A E lR;"xn have singular values UI :::: ... :::: U
r
> 0 and let B E IRfx
q
have singular values <I :::: ... :::: <s > O. Then A 0 B (or B 0 A) has rs singular values
U, <I :::: ... :::: U
r
<s > 0 and
rank(A 0 B) = (rankA)(rankB) = rank(B 0 A) .
n
xn have eigenvalues Ai, i E !!, and let B E IR
m
xm have
eigenvalues JL j, j E m. Then the mn eigenvalues of A 0 Bare
Moreover, if Xl, . , xp are linearly independent right eigenvectors of A corresponding
to AI, ... , A p (p ::::: n), and Z I, ... ,Zq are linearly independent right eigenvectors of B
corresponding to JLI, ... ,JLq (q ::::: m), then Xi 0 Zj E IR
mn
are linearly independent right
eigenvectors of A 0 B corresponding to Ai JL j, i E l!! j E 1
(A 0 B)(x 0 z) = Ax 0 Bz
= AX 0 JLZ
= AJL(X 0 z). 0
If A and Bare diagonalizable in Theorem 13.12, we can take p = nand q = m and
thus get the complete eigenstructure of A 0 B. In general, if A and B have Jordan form
decompositions given by P~
l
AP = JA and Q~
]
BQ = JB, respectively, then we get the
following Jordan-like structure:
Note that JA JB, while upper triangular, is generally not quite in Jordan form and needs
further reduction (to an ultimate Jordan form that also depends on whether or not certain
eigenvalues are zero or nonzero).
A Schur form for A B can be derived similarly. For example, suppose P and
Q are unitary matrices that reduce A and 5, respectively, to Schur (triangular) form, i.e.,
P
H
AP = T
A
and Q
H
BQ = T
B
(and similarly if P and Q are orthogonal similarities
reducing A and B to real Schur form). Then
Corollary 13.13. Let A e R
nxn
and B e R
mxm
. Then
nxn
and B e R
mxm
. Then the Kronecker sum (or tensor sum)
of A and B, denoted A B, is the mn x mn matrix (I
m
< g> A) + (B /). Note that, in
general, A B ^ B A.
Example 13.15.
Then
The reader is invited to compute B 0 A = (/3 B) + (A < g> /2) and note the difference
with A B.
1. Let
142 Chapter 1 3. Kronecker Products
decompositions given by p-
I
AP = J
A
and Q-l BQ = J
B
, respectively, then we get the
following Jordan-like structure:
(P Q)-I(A B)(P Q) = (P-
I
Q-l)(A B)(P Q)
= (P-
1
AP) (Q-l BQ)
= J
A
J
B
Note that h JR, while upper triangular, is generally not quite in Jordan form and needs
further reduction (to an ultimate Jordan form that also depends on whether or not certain
eigenvalues are zero or nonzero).
A Schur form for A B can be derived similarly. For example, suppose P and
Q are unitary matrices that reduce A and B, respectively, to Schur (triangular) form, i.e.,
pH AP = TA and QH BQ = TB (and similarly if P and Q are orthogonal similarities
reducing A and B to real Schur form). Then
(P Q)H (A B)(P Q) = (pH QH)(A B)(P Q)
= (pH AP) (QH BQ)
= TA T
R
.
Corollary 13.13. Let A E IR
n
xn and B E IR
rn
xm. Then
1. Tr(A B) = (TrA)(TrB) = Tr(B A).
2. det(A B) = (det A)m(det Bt = det(B A).
Definition 13.14. Let A E IR
n
Xn and B E IR
m
xrn. Then the Kronecker sum (or tensor sum)
of A and B, denoted A EEl B, is the mn x mn matrix Urn A) + (B In). Note that, in
general, A EEl B i= B EEl A.
Example 13.15.
1. Let

2
;

2
Then
2 3 0 0 0 2 0 0 0 0
3 2 1 0 0 0 0 2 0 0 1 0
AfflB = (hA)+(Bh) =
1 1 4 0 0 0 0 0 2 0 0
0 0 0 2 3
+
2 0 0 3 0 0
0 0 0 3 2 0 2 0 0 3 0
0 0 0 4 0 0 2 0 0 3
The reader is invited to compute B EEl A = (h B) + (A 0 h) and note the difference
with A EEl B.
13.2. Properties of the Kronecker Product 143
If A and B are diagonalizable in Theorem 13.16, we can take p = n and q = m and
thus get the complete eigenstructure of A 0 B. In general, if A and B have Jordan form
decompositions given by P~
1
AP = JA and Q"
1
BQ = JB, respectively, then
is a Jordan-like structure for A B.
Then J can be written in the very compact form J = (4 < 8 > M) + (E^l2) = M 0 Ek.
Theorem 13.16. Let A e E"
x
" have eigenvalues A, - , i e n, and let B e R
mx
'" have
eigenvalues /z
;
, 7 e ra. TTzen r/ze Kronecker sum A B = (I
m
(g> A) + (B < g> /) /za^ ran
e/genva/wes
Moreover, if x\,... ,x
p
are linearly independent right eigenvectors of A corresponding
to AI, . . . , X
p
(p < n), and z\, ..., z
q
are linearly independent right eigenvectors of B
corresponding to f j i \ , . . . , f^
q
(q < ra), then Zj < 8 > Xi W
1
" are linearly independent right
eigenvectors of A B corresponding to A., + [ij , i p, j e q.
2. Recall the real JCF
where M =
13.2. Properties of the Kronecker Product
2. Recall the real JCF
1=
where M = [
a
-f3
M I 0 0
f3
a
o M I 0
M
0
J. Define
0 0
0 0
Ek =
0
I 0
M I
o M
o
o
o
143
E jR2kx2k,
Then 1 can be written in the very compact form 1 = (I} M) + (Ek h) = M $ E
k
.
Theorem 13.16. Let A E jRnxn have eigenvalues Ai, i E !!. and let B E jRmxm have
eigenvalues fJ-j, j E I!!. Then the Kronecker sum A $ B = (1m A) + (B In) has mn
eigenvalues
Al + fJ-t, ... , AI + fJ-m, A2 + fJ-t,, A2 + fJ-m, ... , An + fJ-m'
Moreover, if XI, . , xp are linearly independent right eigenvectors of A corresponding
to AI, ... , Ap (p ::s: n), and ZI, ... , Zq are linearly independent right eigenvectors of B
corresponding to fJ-t, ... , fJ-q (q ::s: m), then Z j Xi E jRmn are linearly independent right
eigenvectors of A $ B corresponding to Ai + fJ-j' i E E, j E fl
[(1m A) + (B In)](Z X) = (Z Ax) + (Bz X)
= (Z Ax) + (fJ-Z X)
= (A + fJ-)(Z X). 0
If A and Bare diagonalizable in Theorem 13.16, we can take p = nand q = m and
thus get the complete eigenstructure of A $ B. In general, if A and B have Jordan form
decompositions given by p-I AP = lA and Q-t BQ = l
B
, respectively, then
[(Q In)(lm p)rt[(lm A) + (B In)][CQ In)(lm P)]
= [(1m p)-I(Q In)-I][(lm A) + (B In)][(Q In)(/m P)]
= [(1m p-I)(Q-I In)][(lm A) + (B In)][CQ In)(/m <:9 P)]
= (1m lA) + (JB In)
is a Jordan-like structure for A $ B.
A Schur form for A B can be derived similarly. Again, suppose P and Q are unitary
matrices that reduce A and B, respectively, to Schur (triangular) form, i.e., P
H
AP = T
A
and Q
H
BQ = T
B
(and similarly if P and Q are orthogonal similarities reducing A and B
to real Schur form). Then
((Q /)(/ P)]"[(/m < 8 > A) + (B /
B
)][(e (g) /)(/, P)] = (/
m
< 8 > r
A
) + (7* (g) /),
where [(Q < 8 > /)(/ P)] = (< 2 P) is unitary by Theorem 13.3 and Corollary 13.8 .
In this section we study the linear matrix equation
A special case of (13.3) is the symmetric equation
obtained by taking B = A
T
. When C is symmetric, the solution X e W
x
" is easily shown
also to be symmetric and (13.4) is known as a Lyapunov equation. Lyapunov equations
arise naturally in stability theory.
The first important question to ask regarding (13.3) is, When does a solution exist?
By writing the matrices in (13.3) in terms of their columns, it is easily seen by equating the
z'th columns that
The coefficient matrix in (13.5) clearly can be written as the Kronecker sum (I
m
* A) +
(B
T
/). The following definition is very helpful in completing the writing of (13.5) as
an "ordinary" linear system.
where A e R"
x
", B e R
mxm
, and C e M"
xm
. This equation is now often called a Sylvester
equation in honor of J.J. Sylvester who studied general linear matrix equations of the form
These equations can then be rewritten as the mn x mn linear system
A Schur fonn for A EB B can be derived similarly. Again, suppose P and Q are unitary
matrices that reduce A and B, respectively, to Schur (triangular) fonn, i.e., pH AP = TA
and QH BQ = TB (and similarly if P and Q are orthogonal similarities reducing A and B
to real Schur fonn). Then
where [(Q In)(lm P)] = (Q P) is unitary by Theorem 13.3 and Corollary 13.8.
In this section we study the linear matrix equation
AX+XB=C, (13.3)
where A E IR
nxn
, B E IR
mxm
, and C E IRnxm. This equation is now often called a Sylvester
equation in honor of 1.1. Sylvester who studied general linear matrix equations of the fonn
k
LA;XB; =C.
;=1
A special case of (13.3) is the symmetric equation
AX +XAT = C (13.4)
obtained by taking B = AT. When C is symmetric, the solution X E IR
n
xn is easily shown
also to be symmetric and (13.4) is known as a Lyapunov equation. Lyapunovequations
arise naturally in stability theory.
The first important question to ask regarding (13.3) is, When does a solution exist?
By writing the matrices in (13.3) in tenns of their columns, it is easily seen by equating the
ith columns that
m
AXi + Xb; = C; = AXi +
j=1
These equations can then be rewritten as the mn x mn linear system
[
A+blll b
21
1
bl21 A + b
2Z
1
blml b2ml
(13.5)
The coefficient matrix in (13.5) clearly can be written as the Kronecker sum (1m 0 A) +
(B
T
0 In). The following definition is very helpful in completing the writing of (13.5) as
an "ordinary" linear system.
13.3. Application to Sylvester and Lyapunov Equations 145
Definition 13.17. Let c
(
E.
n
denote the columns ofC e R
nxm
so that C = [n,..., c
m
}.
Then vec(C) is defined to be the mn-vector formed by stacking the columns ofC on top of
one another, i.e., vec(C) =
Using Definition 13.17, the linear system (13.5) can be rewritten in the form
There exists a unique solution to (13.6) if and only if [(I
m
A) + (B
T
/)] is nonsingular.
But [(I
m
< 8 > A) + (B
T
(g) /)] is nonsingular if and only if it has no zero eigenvalues.
From Theorem 13.16, the eigenvalues of [(/
m
<g> A) + (B
T
<8> /)] are A., + IJ LJ , where
A,,- e A (A), i e n_, and ^j e A(fi), j e m. We thus have the following theorem.
nxn
, B G R
mxm
, and C e R"
xm
. 77ie/i the Sylvester equation
has a unique solution if and only if A and B have no eigenvalues in common.
Sylvester equations of the form (13.3) (or symmetric Lyapunov equations of the form
(13.4)) are generally not solved using the mn x mn "vec" formulation (13.6). The most
commonly preferred numerical algorithm is described in [2]. First A and B are reduced to
(real) Schur form. An equivalent linear system is then solved in which the triangular form
of the reduced A and B can be exploited to solve successively for the columns of a suitably
transformed solution matrix X. Assuming that, say, n > m, this algorithm takes only O(n
3
)
operations rather than the O(n
6
) that would be required by solving (13.6) directly with
Gaussian elimination. A further enhancement to this algorithm is available in [6] whereby
the larger of A or B is initially reduced only to upper Hessenberg rather than triangular
Schur form.
The next few theorems are classical. They culminate in Theorem 13.24, one of many
elegant connections between matrix theory and stability theory for differential equations.
xn
, B e R
mxm
, and C e R
nxm
. Suppose further that A and B
are asymptotically stable (a matrix is asymptotically stable if all its eigenvalues have real
parts in the open left half-plane). Then the (unique) solution of the Sylvester equation
can be written as
Proof: Since A and B are stable, A., (A) + A
;
- (B) ^ 0 for all i, j so there exists a unique
solution to(13.8 )by Theorem 13.18. Now integrate the differential equation X = AX + XB
(with X(0) = C) on [0, +00):
Definition 13.17. Let Ci E jRn denote the columns ofC E jRnxm so that C = [CI, ... , C
m
].
: : ~ ~ : : ~ : : : d ~ ~ : : : O :[]::::fonned by "ocking the colunuu of C on top of
Using Definition 13.17, the linear system (13.5) can be rewritten in the form
[(1m A) + (B
T
In)]vec(X) = vec(C). (13.6)
There exists a unique solution to (13.6) if and only if [(1m A) + (B
T
In)] is nonsingular.
But [(1m A) + (B
T
In)] is nonsingular if and only if it has no zero eigenvalues.
From Theorem 13.16, the eigenvalues of [(1m A) + (BT In)] are Ai + Mj, where
Ai E A(A), i E!!, and Mj E A(B), j E!!!.. We thus have the following theorem.
Theorem 13.1S. Let A E lR
nxn
, B E jRmxm, and C E jRnxm. Then the Sylvester equation
AX+XB=C
(13.7)
has a unique solution if and only if A and - B have no eigenvalues in common.
Sylvester equations of the form (13.3) (or symmetric Lyapunov equations of the form
(13.4 are generally not solved using the mn x mn "vee" formulation (13.6). The most
commonly preferred numerical algorithm is described in [2]. First A and B are reduced to
(real) Schur form. An equivalent linear system is then solved in which the triangular form
of the reduced A and B can be exploited to solve successively for the columns of a suitably
transformed solution matrix X. Assuming that, say, n :::: m, this algorithm takes only 0 (n
3
)
operations rather than the O(n
6
) that would be required by solving (13.6) directly with
Gaussian elimination. A further enhancement to this algorithm is available in [6] whereby
the larger of A or B is initially reduced only to upper Hessenberg rather than triangular
Schur form.
The next few theorems are classical. They culminate in Theorem 13.24, one of many
elegant connections between matrix theory and stability theory for differential equations.
Theorem 13.19. Let A E jRnxn, B E jRmxm, and C E jRnxm. Suppose further that A and B
are asymptotically stable (a matrix is asymptotically stable if all its eigenvalues have real
parts in the open left half-plane). Then the (unique) solution of the Sylvester equation
AX+XB=C (13.8)
can be written as
(13.9)
Proof: Since A and B are stable, Aj(A) + Aj(B) =I 0 for all i, j so there exists a unique
solution to (13.8) by Theorem 13.18. Now integrate the differential equation X = AX + X B
(with X(O) = C) on [0, +00):
lim XU) - X(O) = A roo X(t)dt + ([+00 X(t)dt) B.
I-Hoo 10 10
(13.10)
Using the results of Section 11.1.6, it can be shown easily that lim e = lim e = 0.
r> + oo tv+oo
Hence, using the solution X ( t ) = e
tA
Ce
tB
from Theorem 11.6, we have that lim X ( t ) 0.
/<-+3C
Substituting in (13.10) we have
Remark 13.20. An equivalent condition for the existence of a unique solution to AX +
XB = C is that [ J _
c
fi
] be similar to [ J _
B
] (via the similarity [ J _* ]).
Theorem 13.21. Lef A, C e R"
x
". TTzen r/z e Lyapunov equation
has a unique solution if and only if A and A
T
have no eigenvalues in common. If C is
symmetric and (13.11) has a unique solution, then that solution is symmetric.
Remark 13.22. If the matrix A e W
xn
has eigenvalues A.I ,...,!, then A
T
has eigen-
values A.], . . . , k
n
. Thus, a sufficient condition that guarantees that A and A
T
have
no common eigenvalues is that A be asymptotically stable. Many useful results exist con-
cerning the relationship between stability and Lyapunov equations. Two basic results due
to Lyapunov are the following, the first of which follows immediately from Theorem 13.19.
Theorem 13.23. Let A,C e R"
x
" and suppose further that A is asymptotically stable.
Then the (unique) solution of the Lyapunov equation
can be written as
Theorem 13.24. A matrix A e R"
x
" is asymptotically stable if and only if there exists a
positive definite solution to the Lyapunov equation
Proof: Suppose A is asymptotically stable. By Theorems 13.21 and 13.23 a solution to
(13.13) exists and takes the form (13.12). Now let v be an arbitrary nonz ero vector in E".
Then
and so X
where C -
Using the results of Section 11.1.6, it can be shown easily that lim e
lA
= lim e
lB
= O.
1-->+00 1 .... +00
Hence, using the solution X (t) = elACe
lB
from Theorem 11.6, we have that lim X (t) = O.
t ~ + x
Substituting in (13.10) we have
-C = A (1+
00
elACe
lB
dt) + (1+
00
elACe
lB
dt) B
{+oo
and so X = -1o elACe
lB
dt satisfies (13.8). o
Remark 13.20. An equivalent condition for the existence of a unique solution to AX +
X B = C is that [ ~ _C
B
] be similar to [ ~ _OB] (via the similarity [ ~ _ ~ ]).
Theorem 13.21. Let A, C E jRnxn. Then the Lyapunov equation
AX+XAT = C (13.11)
has a unique solution if and only if A and - A T have no eigenvalues in common. If C is
symmetric and ( 13.11) has a unique solution, then that solution is symmetric.
Remark 13.22. If the matrix A E jRn xn has eigenvalues )"" ... , An, then - AT has eigen-
values -AI, ... , - An. Thus, a sufficient condition that guarantees that A and - A T have
no common eigenvalues is that A be asymptotically stable. Many useful results exist con-
cerning the relationship between stability and Lyapunov equations. Two basic results due
to Lyapunov are the following, the first of which follows immediately from Theorem 13.19.
Theorem 13.23. Let A, C E jRnxn and suppose further that A is asymptotically stable.
Then the (unique) solution o/the Lyapunov equation
AX+XAT=C
can be written as
(13.12)
Theorem 13.24. A matrix A E jRnxn is asymptotically stable if and only if there exists a
positive definite solution to the Lyapunov equation
AX +XAT = C, (13.13)
where C = C
T
< O.
Proof: Suppose A is asymptotically stable. By Theorems l3.21 and l3.23 a solution to
(13.13) exists and takes the form (13.12). Now let v be an arbitrary nonzero vector in jRn.
Then
Since C > 0 and e
tA
is nonsingular for all t, the integrand above is positive. Hence
v
T
Xv > 0 and thus X is positive definite.
Conversely, suppose X = X
T
> 0 and let A. e A (A) with corresponding left eigen-
vector y. Then
Since y
H
Xy > 0, we must have A + A = 2 Re A < 0 . Since A was arbitrary, A must be
asymptotically stable. D
Remark 13.25. The Lyapunov equation AX + XA
T
= C can also be written using the
vec notation in the equivalent form
A subtle point arises when dealing with the "dual" Lyapunov equation A
T
X + XA = C.
The equivalent "vec form" of this equation is
However, the complex-valued equation A
H
X + XA = C is equivalent to
The vec operator has many useful properties, most of which derive from one key
result.
Theorem 13.26. For any three matrices A, B, and C for which the matrix product ABC is
defined,
Proof: The proof follows in a fairly straightforward fashion either directly from the defini-
tions or from the fact that vec(;t;y
r
) = y <8 > x. D
An immediate application is to the derivation of existence and uniqueness conditions
for the solution of the simple Sylvester-like equation introduced in Theorem 6.11.
mxn
, B e R
px(}
, and C e R
mxq
. Then the equation
has a solution X e R.
nxp
if and only ifAA
+
CB
+
B = C, in which case the general solution
is of the form
where Y e R
nxp
is arbitrary. The solution of (13.14) is unique if BB
+
A
+
A = I.
Proof: Write (13.14) as
Since -C > 0 and etA is nonsingular for all t, the integrand above is positive. Hence
v
T
Xv > 0 and thus X is positive definite.
Conversely, suppose X = XT > 0 and let A E A(A) with corresponding left eigen-
vector y. Then
0> yHCy = yH AXy + yHXAT Y
= (A + I)yH Xy.
Since yH Xy > 0, we must have A + I = 2 Re A < O. Since A was arbitrary, A must be
asymptotically stable. D
Remark 13.25. The Lyapunov equation AX + X A T = C can also be written using the
vec notation in the equivalent form
[(/ A) + (A l)]vec(X) = vec(C).
A subtle point arises when dealing with the "dual" Lyapunov equation A T X + X A = C.
The equivalent "vec form" of this equation is
[(/ AT) + (AT l)]vec(X) = vec(C).
However, the complex-valued equation A H X + X A = C is equivalent to
[(/ AH) + (AT l)]vec(X) = vec(C).
The vec operator has many useful properties, most of which derive from one key
result.
Theorem 13.26. For any three matrices A, B, and C for which the matrix product ABC is
defined,
vec(ABC) = (C
T
A)vec(B).
Proof: The proof follows in a fairly straightforward fashion either directly from the defini-
tions or from the fact that vec(xyT) = y x. D
An immediate application is to the derivation of existence and uniqueness conditions
for the solution of the simple Sylvester-like equation introduced in Theorem 6.11.
Theorem 13.27. Let A E jRrnxn, B E jRPxq, and C E jRrnxq. Then the equation
AXB =C (13.14)
has a solution X E jRn x p if and only if A A + C B+ B = C, in which case the general solution
is of the form
(13.15)
where Y E jRnxp is arbitrary. The solution of (13. 14) is unique if BB+ A+ A = [.
Proof: Write (13.14) as
(B
T
A)vec(X) = vec(C) (13.16)
by Theorem 13.26. This "vector equation" has a solution if and only if
It is a straightforward exercise to show that (M N)
+
= M
+
< 8> N
+
. Thus, (13.16) has a
solution if and only if
and hence if and only if AA
+
CB
+
B = C.
The general solution of (13.16) is then given by
where Y is arbitrary. This equation can then be rewritten in the form
or, using Theorem 13.26,
The solution is clearly unique if BB
+
< 8> A
+
A = I. D
EXERCISES
1. For any two matrices A and B for which the indicated matrix product is defined,
show that (vec(A))
r
(vec(fl)) = Tr(A
r
). In particular, if B e Rn
xn
, then Tr(fl) =
vec(/J
r
vec(fl).
2. Prove that for all matrices A and B, (A B)
+
= A
+
B
+
.
3. Show that the equation AX B = C has a solution for all C if A has full row rank and
B has full column rank. Also, show that a solution, if it exists, is unique if A has full
column rank and B has full row rank. What is the solution in this case?
4. Show that the general linear equation
can be written in the form
148 Chapter 1 3. Kronecker Products
by Theorem 13.26. This "vector equation" has a solution if and only if
(B
T
A)(B
T
A)+ vec(C) = vec(C).
It is a straightforward exercise to show that (M N) + = M+ N+. Thus, (13.16) has a
solution if and only if
vec(C) = (B
T
A)B+{ A+)vec(C)
= [(B+ B{ AA+]vec(C)
= vec(AA +C B+ B)
and hence if and only if AA + C B+ B = C.
The general solution of (13 .16) is then given by
vec(X) = (B
T
A) + vec(C) + [I - (B
T
A) + (B
T
A)]vec(Y),
where Y is arbitrary. This equation can then be rewritten in the form
vec(X) = B+{ A+)vec(C) + [I - (BB+{ A+ A]vec(y)
or, using Theorem 13.26,
The solution is clearly unique if B B+ A + A = I. 0
EXERCISES
I. For any two matrices A and B for which the indicated matrix product is defined,
show that (vec(AT (vec(B = Tr(A
T
B). In particular, if B E lR
nxn
, then Tr(B) =
vec(Inl vec(B).
2. Prove that for all matrices A and B, (A B)+ = A+ B+.
3. Show that the equation AX B = C has a solution for all C if A has full row rank and
B has full column rank. Also, show that a solution, if it exists, is unique if A has full
column rank and B has full row rank. What is the solution in this case?
4. Show that the general linear equation
k
LAiXB
i
=C
i=1
can be written in the form
[BT AI + ... + B[ Ak]vec(X) = vec(C).
Exercises 149
5. Let x M
m
and y e E". Show that *
r
< 8 > y = yx
T
.
6. Let A e R"
xn
and e M
mxm
.
(a) Show that ||A < 8 > B||
2
= ||A||
2
||||
2
.
(b) What is ||A B\\
F
in terms of the Frobenius norms of A and B? Justify your
answer carefully.
(c) What is the spectral radius of A < 8 > B in terms of the spectral radii of A and B?
Justify your answer carefully.
7. Let A, 5 eR"
x
".
(a) Show that (/ A)* = / < 8 > A* and (fl < g > /)* = B
fc
/ for all integ ers &.
(b) Show that e
l
A
= I < g ) e
A
and e
5
7
= e
B
(g ) /.
(c) Show that the matrices / (8 ) A and B / commute.
(d) Show that
(Note: This result would look a little "nicer" had we defined our Kronecker
sum the other way around. However, Definition 13.14 is conventional in the
literature.)
8 . Consider the Lyapunov matrix equation (13.11) with
and C the symmetric matrix
Clearly
is a symmetric solution of the equation. Verify that
is also a solution and is nonsymmetric. Explain in lig ht of Theorem 13.21.
9. Block Triangularization: Let
where A e Rn
xn
and D e R
mxm
. It is desired to find a similarity transformation
of the form
such that T
l
ST is block upper triang ular.
Exercises 149
5. Let x E ]Rm and y E ]Rn. Show that x T y = y X T
(a) Show that IIA BII2 = IIAII2I1Blb.
(b) What is II A B II F in terms of the Frobenius norms of A and B? Justify your
answer carefully.
(c) What is the spectral radius of A B in terms of the spectral radii of A and B?
Justify your answer carefully.
7. Let A, B E ]Rnxn.
(a) Show that (l A)k = I Ak and (B Il = Bk I for all integers k.
(b) Show that elA = I e
A
and eB1 = e
B
I.
(c) Show that the matrices I A and B I commute.
(d) Show that
e
AEIlB
= eUA)+(Bl) = e
B
e
A
.
(Note: This result would look a little "nicer" had we defined our Kronecker
sum the other way around. However, Definition 13.14 is conventional in the
literature.)
8. Consider the Lyapunov matrix equation (13.11) with
A = [ ~ _ ~ ]
and C the symmetric matrix
[
Clearly
Xs = [ ~ ~ ]
is a symmetric solution of the equation. Verify that
Xns = [ _ ~ ~ ]
is also a solution and is nonsymmetric. Explain in light of Theorem 13.21.
9. Block Triangularization: Let
where A E ]Rn xn and D E ]Rm xm. It is desired to find a similarity transformation
of the form
T = [ ~ ~ J
such that T-
1
ST is block upper triangular.
(a) Show that S is similar to
if X satisfies the so-called matrix Riccati equation
(b) Formulate a similar result for block lower triangularization of S.
10. Block Diagonalization: Let
where A e Rn
xn
and D E R
mxm
. It is desired to find a similarity transformation of
the form
such that T
l
ST is block diagonal,
if Y satisfies the Sylvester equation
(b) Formulate a similar result for block diagonalization of
[
A +OBX B ]
D-XB
if X satisfies the so-called matrix Riccati equation
C-XA+DX-XBX=O.
(b) Fonnulate a similar result for block lower triangularization of S.
to. Block Diagonalization: Let
S= [ ~ ~ l
where A E jRnxn and D E jRmxm. It is desired to find a similarity transfonnation of
the fonn
T = [ ~ ~ ]
such that T-
1
ST is block diagonal.
if Y satisfies the Sylvester equation
AY - YD = -B.
(b) Fonnulate a similar result for block diagonalization of
Bibliography
[1] Albert, A., Regression and the Moore-Penrose Pseudoinverse, Academic Press, New
York, NY, 1972.
[2] Bartels, R.H., and G.W. Stewart, "Algorithm 432. Solution of the Matrix Equation
AX + XB = C," Cornm. ACM, 15(1972), 820-826.
[3] Bellman, R., Introduction to Matrix Analysis, Second Edition, McGraw-Hill, New
York, NY, 1970.
[4] Bjorck, A., Numerical Methods for Least Squares Problems, SIAM, Philadelphia, PA,
1996.
[5] Cline, R.E., "Note on the Generalized Inverse of the Product of Matrices," SIAM Rev.,
6(1964), 5758.
[6] Golub, G.H., S. Nash, and C. Van Loan, "A Hessenberg-Schur Method for the Problem
AX + XB = C," IEEE Trans. Autom. Control, AC-24(1979), 909-913.
[7] Golub, G.H., and C.F. Van Loan, Matrix Computations, Third Edition, Johns Hopkins
Univ. Press, Baltimore, MD, 1996.
[8] Golub, G.H., and J.H. Wilkinson, "Ill-Conditioned Eigensystems and the Computation
of the Jordan Canonical Form," SIAM Rev., 18(1976), 578-619.
[9] Greville, T.N.E., "Note on the Generalized Inverse of a Matrix Product," SIAM Rev.,
8(1966), 518521 [Erratum, SIAM Rev., 9(1967), 249].
[10] Halmos, PR., Finite-Dimensional Vector Spaces, Second Edition, Van Nostrand,
Princeton, NJ, 1958.
[11] Higham, N.J., Accuracy and Stability of'Numerical Algorithms, Second Edition, SIAM,
Philadelphia, PA, 2002.
[12] Horn, R.A., and C.R. Johnson, Matrix Analysis, Cambridge Univ. Press, Cambridge,
UK, 1985.
[13] Horn, R.A., and C.R. Johnson, Topics in Matrix Analysis, Cambridge Univ. Press,
Cambridge, UK, 1991.
151
Bibliography
[1] Albert, A., Regression and the Moore-Penrose Pseudoinverse, Academic Press, New
York, NY, 1972.
[2] Bartels, RH., and G.w. Stewart, "Algorithm 432. Solution of the Matrix Equation
AX + X B = C," Comm. ACM, 15(1972),820-826.
[3] Bellman, R, Introduction to Matrix Analysis, Second Edition, McGraw-Hill, New
York, NY, 1970.
[4] Bjorck, A., Numerical Methodsfor Least Squares Problems, SIAM, Philadelphia, PA,
1996.
[5] Cline, R.E., "Note on the Generalized Inverse of the Product of Matrices," SIAM Rev.,
6(1964),57-58.
[6] Golub, G.H., S. Nash, and C. Van Loan, "A Hessenberg-Schur Method for the Problem
AX + X B = C," IEEE Trans. Autom. Control, AC-24(1979), 909-913.
[7] Golub, G.H., and c.F. Van Loan, Matrix Computations, Third Edition, Johns Hopkins
Univ. Press, Baltimore, MD, 1996.
[8] Golub, G.H., and lH. Wilkinson, "Ill-Conditioned Eigensystems and the Computation
ofthe Jordan Canonical Form," SIAM Rev., 18(1976),578-619.
[9] Greville, T.N.E., "Note on the Generalized Inverse of a Matrix Product," SIAM Rev.,
8(1966),518-521 [Erratum, SIAM Rev., 9(1967), 249].
[10] Halmos, P.R, Finite-Dimensional Vector Spaces, Second Edition, Van Nostrand,
Princeton, NJ, 1958.
[11] Higham, N.1., Accuracy and Stability of Numerical Algorithms, Second Edition, SIAM,
Philadelphia, PA, 2002.
[12] Hom, RA., and C.R. Johnson, Matrix Analysis, Cambridge Univ. Press, Cambridge,
UK, 1985.
[13] Hom, RA., and C.R. Johnson, Topics in Matrix Analysis, Cambridge Univ. Press,
Cambridge, UK, 1991.
151
152 Bibliography
[14] Kenney, C, and A.J. Laub, "Controllability and Stability Radii for Companion Form
Systems," Math, of Control, Signals, and Systems, 1(1988), 361-390.
[15] Kenney, C.S., and A.J. Laub, "The Matrix Sign Function," IEEE Trans. Autom. Control,
40(1995), 13301348.
[16] Lancaster, P., and M. Tismenetsky, The Theory of Matrices, Second Edition with
Applications, Academic Press, Orlando, FL, 1985.
[17] Laub, A.J., "A Schur Method for Solving Algebraic Riccati Equations," IEEE Trans..
Autom. Control, AC-24( 1979), 913921.
[18] Meyer, C.D., Matrix Analysis and Applied Linear Algebra, SIAM, Philadelphia, PA,
2000.
[19] Moler, C.B., and C.F. Van Loan, "Nineteen Dubious Ways to Compute the Exponential
of a Matrix," SIAM Rev., 20(1978), 801-836.
[20] Noble, B., and J.W. Daniel, Applied Linear Algebra, Third Edition, Prentice-Hall,
Englewood Cliffs, NJ, 1988.
[21] Ortega, J., Matrix Theory. A Second Course, Plenum, New York, NY, 1987.
[22] Penrose, R., "A Generalized Inverse for Matrices," Proc. Cambridge Philos. Soc.,
51(1955), 406413.
[23] Stewart, G. W., Introduction to Matrix Computations, Academic Press, New York, NY,
1973.
[24] Strang, G., Linear Algebra and Its Applications, Third Edition, Harcourt Brace
Jovanovich, San Diego, CA, 1988.
[25] Watkins, D.S., Fundamentals of Matrix Computations, Second Edition, Wiley-
Interscience, New York, 2002.
[26] Wonham, W.M., Linear Multivariable Control. A Geometric Approach, Third Edition,
Springer-Verlag, New York, NY, 1985.
152 Bibliography
[14] Kenney, C., and AJ. Laub, "Controllability and Stability Radii for Companion Fonn
Systems," Math. of Control, Signals, and Systems, 1(1988),361-390.
[15] Kenney, C.S., andAJ. Laub, "The Matrix Sign Function," IEEE Trans. Autom. Control,
40(1995),1330-1348.
[16] Lancaster, P., and M. Tismenetsky, The Theory of Matrices, Second Edition with
Applications, Academic Press, Orlando, FL, 1985.
[17] Laub, AJ., "A Schur Method for Solving Algebraic Riccati Equations," IEEE Trans ..
Autom. Control, AC-24( 1979), 913-921.
[18] Meyer, C.D., Matrix Analysis and Applied Linear Algebra, SIAM, Philadelphia, PA,
2000.
[19] Moler, c.B., and c.P. Van Loan, "Nineteen Dubious Ways to Compute the Exponential
of a Matrix," SIAM Rev., 20(1978),801-836.
[20] Noble, B., and J.w. Daniel, Applied Linear Algebra, Third Edition, Prentice-Hall,
Englewood Cliffs, NJ, 1988.
[21] Ortega, J., Matrix Theory. A Second Course, Plenum, New York, NY, 1987.
[22] Pemose, R., "A Generalized Inverse for Matrices," Proc. Cambridge Philos. Soc.,
51(1955),406-413.
[23] Stewart, G.W., Introduction to Matrix Computations, Academic Press, New York, NY,
1973.
[24] Strang, G., Linear Algebra and Its Applications, Third Edition, Harcourt Brace
Jovanovich, San Diego, CA, 1988.
[25] Watkins, D.S., Fundamentals of Matrix Computations, Second Edition, Wiley-
Interscience, New York, 2002.
[26] Wonham, W.M., Linear Multivariable Control. A Geometric Approach, Third Edition,
Springer-Verlag, New York, NY, 1985.
Index
Ainvariant subspace, 89
matrix characterization of, 90
algebraic multiplicity, 76
angle between vectors, 58
basis, 11
natural, 12
block matrix, 2
definiteness of, 104
diagonalization, 150
inverse of, 48
LU factorization, 5
triangularization, 149
C", 1
(pmxn i
(p/nxn 1
CauchyBunyakovskySchwarz Inequal-
ity, 58
CayleyHamilton Theorem, 75
chain
of eigenvectors, 87
characteristic polynomial
of a matrix, 75
of a matrix pencil, 125
Cholesky factorization, 101
codomain, 17
column
rank, 23
vector, 1
companion matrix
inverse of, 105
pseudoinverse of, 106
singular values of, 106
singular vectors of, 106
complement
of a subspace, 13
orthogonal, 21
congruence, 103
conjugate transpose, 2
contragredient transformation, 137
controllability, 46
defective, 76
degree
of a principal vector, 85
determinant, 4
of a block matrix, 5
properties of, 46
dimension, 12
direct sum
of subspaces, 13
domain, 17
eigenvalue, 75
invariance under similarity transfor-
mation, 81
elementary divisors, 84
equivalence transformation, 95
orthogonal, 95
unitary, 95
equivalent generalized eigenvalue prob-
lems, 127
equivalent matrix pencils, 127
exchange matrix, 39, 89
exponential of a Jordan block, 91, 115
exponential of a matrix, 81, 109
computation of, 114118
inverse of, 110
properties of, 109112
field, 7
four fundamental subspaces, 23
function of a matrix, 81
generalized eigenvalue, 125
generalized real Schur form, 128
153
Index
A-invariant subspace, 89
matrix characterization of, 90
algebraic multiplicity, 76
angle between vectors, 58
basis, 11
natural, 12
block matrix, 2
definiteness of, 104
diagonalization, 150
inverse of, 48
LV factorization, 5
triangularization, 149
en, 1
e
mxn
, 1
e ~ x n , 1
Cauchy-Bunyakovsky-Schwarz Inequal-
ity,58
Cayley-Hamilton Theorem, 75
chain
of eigenvectors, 87
characteristic polynomial
of a matrix, 75
of a matrix pencil, 125
Cholesky factorization, 101
co-domain, 17
column
rank, 23
vector, 1
companion matrix
inverse of, 105
singular vectors of, 106
complement
of a subspace, 13
orthogonal, 21
153
congruence, 103
conjugate transpose, 2
contragredient transformation, 137
controllability, 46
defective, 76
degree
of a principal vector, 85
determinant, 4
properties of, 4-6
dimension, 12
direct sum
of subspaces, 13
domain, 17
eigenvalue, 75
invariance under similarity transfor-
mation,81
elementary divisors, 84
equivalence transformation, 95
orthogonal, 95
unitary, 95
equivalent generalized eigenvalue prob-
lems, 127
equivalent matrix pencils, 127
exchange matrix, 39, 89
exponential of a Jordan block, 91, 115
exponential of a matrix, 81, 109
computation of, 114-118
inverse of, 110
properties of, 109-112
field, 7
four fundamental subspaces, 23
function of a matrix, 81
generalized eigenvalue, 125
generalized real Schur form, 128
154 Index
generalized Schur form, 127
generalized singular value decomposition,
134
geometric multiplicity, 76
Holder Inequality, 58
Hermitian transpose, 2
higherorder difference equations
conversion to firstorder form, 121
higherorder differential equations
higherorder eigenvalue problems
i, 2
idempotent, 6, 51
identity matrix, 4
inertia, 103
initialvalue problem, 109
for higherorder equations, 120
for homogeneous linear difference
equations, 118
for homogeneous linear differential
equations, 112
for inhomogeneous linear difference
equations, 119
for inhomogeneous linear differen-
tial equations, 112
inner product
complex, 55
complex Euclidean, 4
Euclidean, 4, 54
real, 54
usual, 54
weighted, 54
invariant factors, 84
inverses
of block matrices, 47
7, 2
Jordan block, 82
Jordan canonical form (JCF), 82
Kronecker canonical form (KCF), 129
Kronecker delta, 20
Kronecker product, 139
determinant of, 142
eigenvalues of, 141
eigenvectors of, 141
products of, 140
trace of, 142
transpose of, 140
Kronecker sum, 142
eigenvalues of, 143
exponential of, 149
leading principal submatrix, 100
left eigenvector, 75
left generalized eigenvector, 125
left invertible, 26
left nullspace, 22
left principal vector, 85
linear dependence, 10
linear equations
characterization of all solutions, 44
existence of solutions, 44
uniqueness of solutions, 45
linear independence, 10
linear least squares problem, 65
general solution of, 66
geometric solution of, 67
residual of, 65
solution via QR factorization, 71
solution via singular value decom-
position, 70
statement of, 65
uniqueness of solution, 66
linear regression, 67
linear transformation, 17
codomain of, 17
composition of, 19
domain of, 17
invertible, 25
left invertible, 26
matrix representation of, 18
nonsingular, 25
nullspace of, 20
154
generalized Schur form, 127
generalized singular value decomposition,
134
geometric multiplicity, 76
Holder Inequality, 58
Hermitian transpose, 2
higher-order difference equations
conversion to first-order form, 121
higher-order differential equations
higher-order eigenvalue problems
i,2
idempotent, 6, 51
identity matrix, 4
inertia, 103
initial-value problem, 109
for higher-order equations, 120
for homogeneous linear difference
equations, 118
for homogeneous linear differential
equations, 112
for inhomogeneous linear difference
equations, 119
for inhomogeneous linear differen-
tial equations, 112
inner product
complex, 55
complex Euclidean, 4
Euclidean, 4, 54
real, 54
usual, 54
weighted, 54
invariant factors, 84
inverses
of block matrices, 47
j,2
Jordan block, 82
Jordan canonical form (JCF), 82
Kronecker canonical form (KCF), 129
Kronecker delta, 20
Kronecker product, 139
determinant of, 142
eigenvalues of, 141
products of, 140
singUlar values of, 141
trace of, 142
transpose of, 140
Kronecker sum, 142
eigenvalues of, 143
exponential of, 149
leading principal submatrix, 100
left eigenvector, 75
left generalized eigenvector, 125
left invertible. 26
left nullspace, 22
left principal vector, 85
linear dependence, 10
linear equations
Index
characterization of all solutions, 44
existence of solutions, 44
uniqueness of solutions, 45
linear independence, 10
linear least squares problem, 65
general solution of, 66
geometric solution of, 67
residual of, 65
solution via QR factorization, 71
solution via singular value decom-
position, 70
statement of, 65
linear regression, 67
linear transformation, 17
co-domain of, 17
composition of, 19
domain of, 17
invertible, 25
left invertible. 26
matrix representation of, 18
nonsingular, 25
nulls pace of, 20
Index 155
range of, 20
right invertible, 26
LU factorization, 6
block, 5
Lyapunov differential equation, 113
Lyapunov equation, 144
and asymptotic stability, 146
integral form of solution, 146
symmetry of solution, 146
matrix
asymptotically stable, 145
best rank k approximation to, 67
companion, 105
defective, 76
definite, 99
derogatory, 106
diagonal, 2
exponential, 109
Hamiltonian, 122
Hermitian, 2
Householder, 97
indefinite, 99
lower Hessenberg, 2
lower triangular, 2
nearest singular matrix to, 67
nilpotent, 115
nonderogatory, 105
normal, 33, 95
orthogonal, 4
pentadiagonal, 2
quasiuppertriangular, 98
sign of a, 91
square root of a, 101
symmetric, 2
symplectic, 122
tridiagonal, 2
unitary, 4
upper Hessenberg, 2
upper triangular, 2
matrix exponential, 81, 91, 109
matrix norm, 59
1.60
2, 60
oo, 60
/?, 60
consistent, 61
Frobenius, 60
induced by a vector norm, 61
mixed, 60
mutually consistent, 61
relations among, 61
Schatten, 60
spectral, 60
subordinate to a vector norm, 61
unitarily invariant, 62
matrix pencil, 125
equivalent, 127
reciprocal, 126
regular, 126
singular, 126
matrix sign function, 91
minimal polynomial, 76
monic polynomial, 76
MoorePenrose pseudoinverse, 29
multiplication
matrixmatrix, 3
matrixvector, 3
MurnaghanWintner Theorem, 98
negative definite, 99
negative invariant subspace, 92
nonnegative definite, 99
criteria for, 100
nonpositive definite, 99
norm
induced, 56
natural, 56
normal equations, 65
normed linear space, 57
nullity, 24
nullspace, 20
left, 22
right, 22
observability, 46
onetoone (11), 23
conditions for, 25
onto, 23
conditions for, 25
Index
range of, 20
LV factorization, 6
block,5
Lyapunov differential equation, 113
Lyapunov equation, 144
and asymptotic stability, 146
symmetry of solution, 146
matrix
asymptotically stable, 145
best rank k approximation to, 67
companion, 105
defective, 76
definite, 99
derogatory, 106
diagonal,2
exponential, 109
Hamiltonian, 122
Hermitian, 2
Householder, 97
indefinite, 99
lower Hessenberg, 2
lower triangular, 2
nearest singular matrix to, 67
nilpotent, 115
nonderogatory, 105
normal, 33, 95
orthogonal, 4
pentadiagonal, 2
quasi-upper-triangular, 98
sign of a, 91
square root of a, 10 1
symmetric, 2
symplectic, 122
tridiagonal, 2
unitary, 4
upper Hessenberg, 2
upper triangular, 2
matrix exponential, 81, 91, 109
matrix norm, 59
1-,60
2-,60
00-,60
p-,60
consistent, 61
Frobenius, 60
induced by a vector norm, 61
mixed,60
mutually consistent, 61
relations among, 61
Schatten,60
spectral, 60
155
subordinate to a vector norm, 61
matrix pencil, 125
equivalent, 127
reciprocal, 126
regular, 126
singUlar, 126
matrix sign function, 91
minimal polynomial, 76
monic polynomial, 76
Moore-Penrose pseudoinverse, 29
multiplication
matrix-matrix, 3
matrix-vector, 3
Mumaghan-Wintner Theorem, 98
negative definite, 99
negative invariant subspace, 92
nonnegative definite, 99
criteria for, 100
nonpositive definite, 99
norm
induced,56
natural,56
normal equations, 65
normed linear space, 57
nullity, 24
nullspace,20
left, 22
right, 22
observability, 46
one-to-one (1-1), 23
conditions for, 25
onto, 23
conditions for, 25
156 Index
orthogonal
complement, 21
matrix, 4
projection, 52
subspaces, 14
vectors, 4, 20
orthonormal
vectors, 4, 20
outer product, 19
and Kronecker product, 140
exponential of, 121
singular value decomposition of, 41
various matrix norms of, 63
pencil
equivalent, 127
of matrices, 125
reciprocal, 126
regular, 126
singular, 126
Penrose theorem, 30
polar factorization, 41
polarization identity, 57
positive definite, 99
criteria for, 100
positive invariant subspace, 92
power (Kth) of a Jordan block, 120
powers of a matrix
computation of, 119120
principal submatrix, 100
projection
oblique, 51
on four fundamental subspaces, 52
orthogonal, 52
pseudoinverse, 29
four Penrose conditions for, 30
of a fullcolumnrank matrix, 30
of a fullrowrank matrix, 30
of a matrix product, 32
of a scalar, 31
of a vector, 31
uniqueness, 30
via singular value decomposition, 38
Pythagorean Identity, 59
Q orthogonality, 55
QR factorization, 72
T O " 1
IK , 1
M
mxn i
, 1
M
mxn 1
r '
M nxn 1
n ' '
range, 20
range inclusion
characterized by pseudoinverses, 33
rank, 23
column, 23
row, 23
rankone matrix, 19
rational canonical form, 104
Rayleigh quotient, 100
reachability, 46
real Schur canonical form, 98
real Schur form, 98
reciprocal matrix pencil, 126
reconstructibility, 46
regular matrix pencil, 126
residual, 65
resolvent, 111
reverseorder identity matrix, 39, 89
right eigenvector, 75
right generalized eigenvector, 125
right nullspace, 22
right principal vector, 85
row
rank, 23
vector, 1
Schur canonical form, 98
generalized, 127
Schur complement, 6, 48, 102, 104
Schur T heorem, 98
Schur vectors, 98
secondorder eigenvalue problem, 135
ShermanMorrisonWoodbury formula,
48
signature, 103
similarity transformation, 95
and invariance of eigenvalues, h
156
orthogonal
complement, 21
matrix, 4
projection, 52
subspaces, 14
vectors, 4, 20
orthonormal
vectors, 4, 20
outer product, 19
and Kronecker product, 140
exponential of, 121
singular value decomposition of, 41
various matrix norms of, 63
pencil
equivalent, 127
of matrices, 125
reciprocal, 126
regular, 126
singular, 126
Penrose theorem, 30
polar factorization, 41
polarization identity, 57
positive definite, 99
criteria for, 100
positive invariant subspace, 92
power (kth) of a Jordan block, 120
powers of a matrix
computation of, 119-120
principal submatrix, 100
projection
oblique, 51
on four fundamental subspaces, 52
orthogonal, 52
pseudoinverse, 29
four Penrose conditions for, 30
of a full-column-rank matrix, 30
of a full-row-rank matrix, 30
of a scalar, 31
of a vector, 31
uniqueness, 30
Pythagorean Identity, 59
Q-orthogonality, 55
QR factorization, 72
JR.n, I
JR.mxn,1
1
I
range, 20
range inclusion
Index
characterized by pseudoinverses, 33
rank, 23
column, 23
row, 23
rank-one matrix, 19
rational canonical form, 104
Rayleigh quotient, 100
reachability, 46
real Schur canonical form, 98
real Schur form, 98
reciprocal matrix pencil, 126
reconstructibility, 46
regular matrix pencil, 126
residual, 65
resolvent, III
reverse-order identity matrix, 39, 89
right eigenvector, 75
right generalized eigenvector, 125
right nullspace, 22
right principal vector, 85
row
rank, 23
vector, I
Schur canonical form, 98
generalized, 127
Schur complement, 6, 48, 102, 104
Schur Theorem, 98
Schur vectors, 98
second-order eigenvalue problem, 135
Sherman-Morrison-Woodbury formula,
48
signature, 103
similarity transformation, 95
and invariance of eigenvalues, 81
Index 157
orthogonal, 95
unitary, 95
simple eigenvalue, 85
simultaneous diagonalization, 133
singular matrix pencil, 126
singular value decomposition (SVD), 35
and bases for four fundamental
subspaces, 38
and pseudoinverse, 38
and rank, 38
characterization of a matrix factor-
ization as, 37
dyadic expansion, 38
examples, 37
full vs. compact, 37
fundamental theorem, 35
nonuniqueness, 36
singular values, 36
singular vectors
left, 36
right, 36
span, 11
spectral radius, 62, 107
spectral representation, 97
spectrum, 76
subordinate norm, 61
subspace, 9
Ainvariant, 89
deflating, 129
reducing, 130
subspaces
complements of, 13
direct sum of, 13
equality of, 10
four fundamental, 23
intersection of, 13
orthogonal, 14
sum of, 13
Sylvester differential equation, 113
Sylvester equation, 144
Sylvester's Law of Inertia, 103
symmetric generalized eigenvalue prob-
lem, 131
total least squares, 68
trace, 6
transpose, 2
characterization by inner product, 54
triangle inequality
for matrix norms, 59
for vector norms, 57
unitarily invariant
matrix norm, 62
vector norm, 58
variation of parameters, 112
vec
of a matrix, 145
vector norm, 57
l, 57
2, 57
oo, 57
P, 51
equivalent, 59
Euclidean, 57
Manhattan, 57
relations among, 59
weighted, 58
weighted p, 58
vector space, 8
dimension of, 12
vectors, 1
column, 1
linearly dependent, 10
linearly independent, 10
orthogonal, 4, 20
orthonormal, 4, 20
row, 1
span of a set of, 11
zeros
of a linear dynamical system, 130
Index
orthogonal, 95
unitary, 95
simple eigenvalue, 85
simultaneous diagonalization, 133
singular matrix pencil, 126
singular value decomposition (SVD), 35
and bases for four fundamental
subspaces, 38
and pseudoinverse, 38
and rank, 38
characterization of a matrix factor-
ization as, 37
dyadic expansion, 38
examples, 37
full vs. compact, 37
fundamental theorem, 35
nonuniqueness, 36
singular values, 36
singular vectors
left, 36
right, 36
span, 11
spectral radius, 62, 107
spectral representation, 97
spectrum, 76
subordinate norm, 61
subspace, 9
A-invariant, 89
deflating, 129
reducing, 130
subspaces
complements of, 13
direct sum of, 13
equality of, 10
four fundamental, 23
intersection of, 13
orthogonal, 14
sum of, 13
Sylvester differential equation, 113
Sylvester equation, 144
157
Sylvester's Law of Inertia, 103
symmetric generalized eigenvalue prob-
lem,131
total least squares, 68
trace, 6
transpose, 2
characterization by inner product, 54
triangle inequality
for matrix norms, 59
for vector norms, 57
unitarily invariant
matrix norm, 62
vector norm, 58
variation of parameters, 112
vec
of a matrix, 145
vector norm, 57
1-,57
2-,57
00-,57
p-,57
equivalent, 59
Euclidean, 57
Manhattan, 57
relations among, 59
weighted, 58
weighted p-, 58
vector space, 8
dimension of, 12
vectors, 1
column, 1
linearly dependent, 10
linearly independent, 10
orthogonal, 4, 20
orthonormal, 4, 20
row, 1
span of a set of, 11
zeros
of a linear dynamical system, 130

Matrix Analysis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Matrix Analysis

Uploaded by

Copyright:

Available Formats

Matrix Analysis

for Scientists & Engineers

and 7(A). The transformations A

, then T can be defined by TV; = cr, w, , / e r. Similarly, since [u\, ... ,u

A(A) I ' M- Then p(A) is called the spectral

You might also like